Something is hilariously off here: Why should I pay $10 and be forced to use it by the end of the month, while I can pay $10 and have it last as long as I want?
How would that be? They are already charging as much as the underlying providers. They can hardly expect to have any customers if they are charging more.
Enterprise sales will be the answer. Microsoft will have some story that convinces an exec eight levels up the org chart from the normal users that this is an essential product they need to overpay for. Given their existing relationshipsand immense sales team they'll probably have success.
That story is data governance. Corporate already have a data-agreement with MS, storing all their data there. Github copilot is covered by that, while a individual agreement with e.g. anthropic needs lawers involved.
It’s precisely this, and, to be fair, it’s a rational approach given a Data Security Exhibit starts at 6 weeks and can hit 6 months to complete. That being said, I work with regulated data, so YMMV.
Plus if you're in government you have procurement to deal with. You already have an Enterprise deal with Microsoft so you don't have to go through any of that rigmarole.
Microsoft is simply the default answer for most large corporations. Getting access to some Microsoft subscription is very easy, because of the existing framework agreements, Microsoft providing any and all compliance slopuments needed and already being pre-cleared for corporate data etc. Meanwhile trying to use another provider (e.g. Anthropic) would be a one year endeavor, minimum.
They list the price 900% higher and give a 90% discount to enterprises who also use teams, outlook, office or even windows if they're desperate. Then that becomes a deal so good that enterprises can't afford not to take it!
e.g. if on an annual plan? 0x will be gone, but there are okay 1x and 0.3x models left. I am pretty much curious how the early may test invoicing will look like. current setup of tools etc. is way too chatty eats up 1+M token per PRU easily. not sure how much is cached.
Only reason to keep it is if you like their UX and auto-complete. Everything else is on pay per use and if you don't use all of it (good luck with the 5 hour and week caps) you have just paid more for the auto-complete
The deal is really pretty much garbage now and I believe that is the intent.
Can you elaborate please? I still clone my private repos and work on them using OpenRouter Credits and OpenCode (or Copilot itself: It supports OpenRouter BYOK). No?
is $10 Pro monthly subscription a pre-requisite before i can purchase $10 in API credits?
PS: i would have loved if I can directly buy $10 in credits and be free to spend it as quickly or as leisurly as I want -- without any monthly expiry or fixed recurring payments
I have to wonder if it's because of how many Enterprise customers they have who have standardized on Github Copilot and gotten it through the gauntlet of legal approvals etc.
This rings true to me as someone who's worked at a few large corps like this. A price hike does not change things when there is a mandate to use MS products over other vendors.
We're putting other providers through the gauntlet. An M4 Studio or two running the latest Qwen3 or whatever counts for state of the art in open models is also looking a little more viable all the time.
They can be super complementary! Open weight models can be your everyday standard goto, and frontier models for the harder and bigger tasks.
Having some open weight deployment or vendor is also a good thing, because you may have domain specific tasks where you can get better results on domain specific problems with a quick finetune.
Unsloth makes it particularly easy. Open weight LLMs are incredibly powerful building blocks.
Bingo. Ghcp is the only allowed llm solution at our big well known semiconductor corp. It'll take years to get approval for anything else. We're stuck with it and will pay whatever price we have to.
I'm wondering if they're basically saying they're going to give $10/month free API credits to students and open source maintainers and so on... while otherwise getting out of the consumer portion of this space.
they're downsizing free github copilot pro for open source maintainers. At the very least, it looks like small open source projects got their free copilot pro cut off
I was kinda of disappointed when that happened earlier this month, but not as much now after seeing this change. My primary use had been trying out some of the newer Anthropic and OpenAI models, which probably would have burned through $10 worth of credits rather quickly given their new pricing.
I am a bit confused by the separation between VSCode and Copilot. If I cancel my Pro+ subscription, can I still use Copilot with my own OpenRouter key?
Huh, I find my copilot plugin to be so incredibly glitchy. My agents are always reporting that their shells are mangled that commands are truncated and all kinds of nonsense. Sometimes they spin up dev servers fine other times it just hangs waiting for a terminal response. So far I have found relying on the CLI from the model providers to be significantly more reliable.
I do like the integrations with the IDE however, they are convenient for rapidly reviewing changes. I just need their terminals to actually work!
I had this problem and it turns out it was my oh-my-posh command prompt customization. VS Code injects certain control characters into the output stream for agents to observe events and the theming runs after those mechanics are hooked up so it can interfere. Updating to the latest oh-my-posh fixed it for me.
Here's the oh-my-posh GH issue[0] in case your problem is similar but not solvable with a simple package update.
Ive also had huge problems with copilot and terminals, mostly due to corporate antivirus which makes it impossible to install clink for example. So the Shell integration never really worked. It has been fixed for a couple of months though now, and its working, maybe not great but at least very good.
The only sad thing is trying to use tools in a VS developer prompt (and how could this not have been fixed ages ago, its literall YOUR OWN flagship product). It knows how to launch the .cmd for it, but thats incredibly slow for single commands. Would be nice if I could tell it to just use an open terminal.
You can pool credits through open router (afaik, I'm only using a single user account), but if you top-up $10 per user, per month, any unused credits will rollover.
Tbh I think it still works, but only because the new allowance will likely get used very quickly within a billing cycle - I'm expecting this change to increase our orgs bill significantly based on how many API credits with open router I consume in a weekend using a single agent in a pairing style.
The pooling will only be useful if you have a bunch of infrequent/low usage users that you still want to have licenses.
Which is almost guaranteed to be the case for a large org, considering everyone will want auto complete and PR reviews, but on average most will not be making a ton of agent use
Lots of us have noticed that usage limits for Claude have been nerfed in recent weeks/months.
If anything, these new multipliers are more transparent than anything OpenAI or Anthropic have communicated regarding actual costs and give us a more realistic understanding of what it's costing these providers.
The fact that we were able to get such a substantial amount of usage for $20/$100/$200 a month was never meant to last and to think otherwise was perhaps a bit naive.
This feels like a strategy from the ZIRP era of tech growth where companies burned investor capital and gave away their products and services for free (or subsidized them heavily) in order to prioritize user acquisition initially. Then once they'd gained enough traction and stickiness they'd then implement a monetization strategy to capitalize on said user base.
However, inference costs for entirely good enough models are likely to keep declining in the future. We're probably hitting diminishing returns on model size and training. The new generations aren't quantum leaps anymore, and newer generations of open source models like DeepSeek are likely to start getting good enough.
There's going to be a limit to how much they can raise prices, because someone can always build out a datacenter and fill it up with open source DeepSeek inference and undercut your prices by 10x while still making a very good ROI--and that's a business model right there. Right now I'm sure there's a lot of people who will protest that they couldn't do their jobs with lesser models, but as time goes on that will get less and less. Already right now the consumers who are using AI for writing presentations, cooking recipe generation and ELI5 answers for common things, aren't going to be missing much from a lesser model. That'll actually only start to get cheaper over time.
Also for business needs, as AI inference costs escalate there comes a point where businesses rediscover human intelligence again, and start hiring/training people to do more work to use lesser models--if that is more productive in the end than shelling out large amounts of cash for inference on the latest models. [Although given how much companies waste on AWS, there's a lot of tolerance for overspending in corporations...]
> because someone can always build out a datacenter and fill it up with open source DeepSeek inference and undercut your prices by 10x while still making a very good ROI-
Not sure how it all works out. Currently trillion dollar companies can't make a native app for platforms. Everything is just JS/Electron because economics does not work for them.
And here companies can make GW data center running very expensive GPUs for 1/10th of current prices. Sound little fanciful to me.
The price you pay for anthropic must include the price of training new and better models which is incredibly costly. If you use the models someone else already spend money to develop you don’t need to pay this price.
They've been like that for a while actually, I think at least since the big hype around ChatGPT 4.5 (or was it 5?) and that underwhelming, lukewarm, oversanitised presentation by Altman and his team.
Yups... Mythos is the smallest possible leap. Not a standard model generation advance, not even a version point advance. Just the smallest possible quanta of a change. We are absolutely hitting a plateau any day now. Any day. Any time. Any second now. Yup. Right now! Surely!
I mean let's be realistic - all that we know about the "mythical" Mythos is the carefully curated and release stuff by the Anthropic's PR team. Is it really a huge leap they are making it to be? I doubt it. In fact I bet if it was indeed that powerful and dangerous, as they imply, they'd find a way to release it immediately, devastate OpenAI and DeepSeek and secure a leading position in the market. Why is it not happening? I suspect because Dario is again at it, peddling his bullshit.
Yeah. AI progress is insanely fast if you compare it to anything else. Where else is a one year old technology already hopelessly outdated? 10 years ago is basically stone age.
I am continually tripped out by the fact when I was 16, I didn't have a 'smartphone' beyond a Windows Mobile 6 phone that had no internet on it.
Now, I have this high-resolution shiny object that can near instantaneously get any information I want along with _streaming HD video to it_ *anywhere*.
15 years even feels like a stone age. I can't fathom what it has to feel like people in their 60s and 70s.
I'm not quite 60, but it's always interesting to me that I feel quite the opposite of this. When I was 16, I didn't have a computer, didn't have a phone, had never used the Internet, but when I think of how life has changed, it's frankly not much. I woke up this morning, scooped my cats' litter boxes, took out some trash, made myself breakfast, ate that, read some news while eating, then lifted weights in my garage, had some work meetings, wrote up some instructions per a customer request from Friday, and am about to go drive to the lake to go do a 9 mile longboard loop.
That's very close to a normal day in 1996. The biggest difference is I read the news on my phone instead of a physical newspaper. The news was not any more interesting or informative because of that. I guess I can also still do the loop reasonably well, but I'm a lot slower than I was in 1996 when I was a cross-country state champion.
My parents are closing in on 70 and I guess I can't speak for them, but I'm at least aware of the daily routines of their lives, too. Walk the dog, do housework, DIY building projects, visit kids and grankids. Seems much the same, too, with the biggest difference being they're now teaching my sister's sons to play baseball rather than me, but shit, one of her sons even looks like exactly the same way I looked when I was 7! The more things change, the more they stay the same.
General agree... I still do the things (mid-50's) I used to do when I was a teenager with no computer, no phone.
But - now they are easier - I can read books on an e-ink screen and pretty much instantly find what I want to read next. I get my news on a phone. I used to watch TV/movies broadcast or on tape rentals. Now, I have just about everything I could ever want available - without ADs... those were such a time-waster.
What has changed is that I have access to MORE information than my local (or school) libraries could ever provide - in a variety of more accessible formats. Whatever tools I need to get "work done", I can find a myriad of free and open-source options.
But - the overall days and household family routines are the same - now, instead of reading a paper book while waiting to pickup my kids (or other family members) "back-in-the-day", I can read my device, or connect with my DIY communities online on my phone - or learn something new. I don't have to schedule life around major broadcast events, I can easily do many tasks while I am "out-and-about".
If your parents are closing in on 70, I would have expected you to be closer to not quite 50 than not quite 60.
I am just over 50 myself and I agree with your points. Technology has changed but life is largely very similar to wear it was in the 90s. At least day to day. Attitudes are way worse now.
I always wonder the views of older people. My parents are very technology forward and have been my entire life so it is difficult to gauge how different life is compared to when they were growing up.
It's easy to hear "Oh well I only had 640kb of memory and typed programs out of a magazine I got in the mail!" and see as distinct from having 'unlimited' resources and the internet.
Your insight is good ("The biggest difference is I read the news on my phone instead of a physical newspaper") that life sort of stays the same but the modality changes. People still go to the store like they did in the mid-1800s but now it is by car.
I wonder what our "industrial revolution" will be where the previous generation lived (ie: out in the country on a farm) totally different lives to the current (ie: in the city in a factory). Maybe when space travel and multi-planetary living is normalized?
> It's easy to hear "Oh well I only had 640kb of memory and typed programs out of a magazine I got in the mail!"
Since I was there (young, but there), I want to point out that this crosses three eras which all felt quite different:
1978: typed programs in from a magazine or loaded from a cassette (16kB, TRS-80)
1983: loaded programs from a floppy (64kB, Apple ][ and C64 etc)
1988: loaded programs from a hard disk (640kB, IBM PC and Mac).
Exact years vary but these eras were only about 5 years each. Nobody had a floppy in 1978 but almost computer user did by 1983; nobody had a hard drive in 1983 but almost everyone did by 1988.
To some degree this already happened with the move from the industrial city to suburbanization and then re-urbanization. In particular one of the most notable recent developments is that urban waterways are now pretty desirable places to be with parks and recreation; in most industrializing cities the waterfront was actively avoided because the industrial use made it polluted, smelly etc.
Depends on where you live. My dad is almost 80, grew up in a very rural area, and when he was 16 they'd just gotten indoor plumbing. Up until he was 14, his school was a one-room school house with no heating other than a wood stove. If you were the first kid to arrive for the day, it was your job to get the fire going in winter months.
Housework meant no laundry machine, no dishwasher, and possibly no vacuum cleaner. That means hand washing everything, and beating rugs with sticks and brushes to get the dust off of them.
And at some point even frontier model costs will hopefully come down (if there is still a meaningful difference between closed and open source models at that point) as all of the compute that's being built out right now comes online.
But the prices haven't been going up by multiples of 6 for the past few years. Things are actually changing now. I don't think it's over, but in the short term, it's going to be considerably more expensive.
They will smooth up the spike. Or be subtle and transform the existing quota so that they run out more quickly. Calling it caching, compression, optimisation, of course for the sacred benefit of the users.
Dunno, if in this day and age you are making inference more expensive, more scarce, you are honestly moving in the wrong direction and DeepSeek and others will gladly take your lunch.
> The hardware to run deepseek is still incredibly expensive.
Deepseek API pricing is very low compared to Anthropic/OpenAI API pricing.
For many, the 300% difference in pricing may be difficult to justify, if the quality difference is very small. And there will be many tasks where the most expensive/the best model, is not needed. Currently many people end up using Opus 4.7/GPT 5.5 for many tasks without thinking about it.
Near zero probability of that. The model is more efficient and the company who trained it did not blunder trillions of dollars to do so. China has better electricity infrastructure than the US too, so the likelihood they can scale out before the US ever could is high. Long term deepseek, Alibaba, etc hold the most cards for sustainable AI even despite the attempted Nvidia embargo
I am not shilling China, this is just what is happening right now.
I think the Chinese government works differently than the US government. I think China has been subsidizing their electricity grid for decades and leading the world on sustainable electricity namely solar. While the us has let their infrastructure rot and laughed at government inefficiencies for about half that time. The US has data centers running on gas right now while waging wars blowing up gas infrastructure world wide. It would be comical if it wasn't an environmental disaster. Most of them have no hopes at even getting enough power in well established areas short term.
I realize what I am saying may come off as propaganda because the US holds net negative views on China so here are some links.
I think because openai spent so much money upfront showing how it was possible to do this and laid out a product roadmap China got to get on board much cheaper and easier. I see no reason to not believe any of these companies when they say they didn't squander tons of money to do what they did because I don't know how openai has even spent all the money they have it's actually ridiculous to think about.
It's not really about that. China is eating the US's lunch when it comes to ai. Don't get me wrong opus is the strongest model out there today, but that's the us's only advantage right now. Deepseek,qwen,kimi, etc all have fundamental research making the models smaller, more efficient, scalable, etc. in the US the plan is to buy all the hardware, write legislature, embargo other countries, keep models and research closed, so people cannot innovate for the next two to five years.
Unlike the us chinas focus is on research and sustainable building. China also has really good infrastructure for energy, etc. it is also to their advantage to drop 5 billion instead of 2 trillion and beat the us while turning a profit.
Chinas focus in ai is less flashy and because they are the biggest manufacturing super power in the world right now, it directly feeds their economy. They aren't looking for applications or to replace thought workers with slop bots, they have natural needs for this technology. Us manufacturers can't compete so they have to keep companies from selling their goods there see byd. China sees it as commoditizing their complement, the us is risking its entire economy and it's environment and resources, kind of scary.
If/when it gets to the point where it can replace a skilled worker, the service can be sold for close to the same price as that skilled labour. But the AI can run 24/7, reliably, and scale up/down at a moments notice.
There's not going to be much competition to drive prices down, the barriers to entry are already huge. There'll likely to be one clear winner, becoming a near-monopoly, or maybe we'll get a duopoly at best.
Yes, a lot of people (not me). Why? Well because that was the whole value proposition of these companies, relentlessly pushed by their PR and most of the media- rememmber it was something something Pocket PhDs, massive unemployment etc?
"There's not going to be much competition to drive prices down, the barriers to entry are already huge. There'll likely to be one clear winner, becoming a near-monopoly, or maybe we'll get a duopoly at best."
Based on what exactly? So far every time OpenAI, Anthropic or whatever has released a new top performing model, competitors have caught up quickly. Open source models have greatly improved as well.
I expect AI to be just like cloud computing in general - AWS, Azure, GCP being the main providers, with dozens of smaller competitors offering similar services as well.
Right now China is flexing the future in my opinion. Smaller, widely available, frontier models for pennies on the dollar.
I think the future of ai will be breakthroughs that let it run on commodity hardware, and the average person will not be paying for it from the cloud unless they want to be surveilled or are stuck on older hardware.
Right now I am running about what was a frontier model 1-2 years ago on a junk machine. Some people are running what was a frontier model 4 months ago on PCs and laptops that cost 5,000. In a year I think the landscape will be even better.
I do. "Commoditize your complement". Want to sell lots of silicon? Give away good local models to run on that silicon.
Even if SOTA models in the cloud are a few percentage points better, most work can be routed to local models most of the time. That leaves the cloud providers fighting over the most computationally intensive tasks. In the long term, I think models are going to be local-first.
(Unless providers can figure out a network effect that local models can't replicate).
Why on earth would that happen when everything else is moving into the cloud to tie it to ever-escalating subscription fees and prevent piracy?
Even with gaming, where running high-end 3D games in the cloud seems like madness and inevitably degrades the quality of the experience, they won't stop trying.
> In the long term, I think models are going to be local-first.
Why? There's an inherent efficiency advantage to scale, while the only real advantage for local models (privacy/secrecy) hasn't proven convincing for broader IT either.
It's foolish not to care about privacy especially as a company. You know how it prevents you from emailing yourself your tax documents? Meanwhile thousands of employees are sending literal design docs, software, product goals, etc to several ai third partys. Not only is that insane, the companies they are sending it too intend too and openly admit to scanning the data, make software products themselves, and intend to create models that can produce their products automatically.
The reason local models hasn't caught on is several fold. It's marketing to say your company follows the latest trend, and there's an inherent pressure to keep AI companies afloat so the economy doesn't entirely collapse. The other is, it wasn't until the last month that these models have caught up to frontier models. They just did, and they are more efficient and don't require a team of 500 to deploy.
Local first models aren't just more private than the API vendors, they also have the advantages of fixed cost, lower latency, and better stability - local models don't get nerfed/"updated" in the background like chatgpt does.
Maybe in a world where these AI companies behaved with some semblance of ethics and user-friendliness they would be on even ground, but for anyone paying attention local models are obviously the future.
That's a silly reason. For non-agent use cases what kind of utilization are you going to average on your own GPU, 5-10%? And that's without batching.
Even with overhead and scaling for peak use and a large profit margin, any company with an ounce of competition will be vastly cheaper than self-hosting. And for models you can run yourself, there will be plenty of competition.
"This change aligns Copilot pricing with actual usage and is an important step toward a sustainable, reliable Copilot business and experience for all users."
I see statements like this as strong indicators that the sales people are wrapping up their work and the accountants are taking over. The land rush is switching to an operational efficiency play.
Just be aware OpenRouter charges a 5.5% fee, I didn’t know until recently. I like the product, and I think the fee is fair, but if you want the absolute best pricing then go direct.
But with open router you can always just use the latest model. If you're committed to eg Claude opus then you're better off going directly to anthropic for sure, but if not, varying other models may be fine too, depending on use case and be massively cheaper. Eg new deep seek model with same mio context window or Kimi k2.6 with 270k context window for subagents which implement
>but if not, varying other models may be fine too, depending on use case and be massively cheaper
Do inference providers have standardized endpoints, or at least endpoints compatible with claude code? Otherwise to pay 5.5% on all your tokens just so it's slightly easier to swap providers (ie. changing a few urls?)
There's nothing trivial about getting a Google API key. Openrouter removes that stress from my life. And I can route requests to providers above a certain TPS threshold. And much more.
I had copilot mainly so I could write issues and throw agents at it, while I went off and did other things. Has been great for contained spot work.
At this point, I'll go ahead and leave it expire, and then consolidate between Codex and JetBrains AI. Especially since Xcode supports Codex with a first-party integration.
They do for any new plan.
Those multipliers are only for people that paid annually. After their subscription ends they'll go into token based pricing like the rest of people.
I understand it like : the 10 usd is for handling the business record, maybe also the harness, I get a few coins to kick tires, but to use it for anything real it’s pay as you go by the tokens list price.
I thought I was smart for buying the annual plan after I graduated and lost my student plan and then GitHub taking away my Copilot Pro I got for free for being a author of a popular OSS project. Turns out I'm being punished for making that year commitment to them. I like to think I'm only a moderate user of GHCP so this is just terrible for me. I'm honestly thinking about cancelling and switching to alternatives while also looking at investing in a local LLM setup.
27x for Opus is genuinely shocking. at that point you're not paying for convenience anymore, you're just paying a GitHub tax. OpenRouter or direct API makes way more sense unless you're really glued to the IDE integration.
Does it effectively bypass regional restrictions for you, so you can use something like the Claude API from unsupported regions such as Hong Kong, or does it still enforce the official providers' geo-restrictions?
OpenRouter is great for budget control, but as they are indirect APIs, your experience with cached tokens may vary, eventually costing much more than in direct depending on the providers.
You can pay with crypto though, which seems to be convenient for people under sanctions or with limited access, or if you are in low-tax jurisdiction (e.g. HK)
That said I think few people using openrouter are actually being selective about providers.
It took half a day to get my opencode setup, was not friendly. A lot of manually cross referencing model and providers. I was actually mainly optimizing for relatively fast providers. It all is super fragile and I'm sure half out of date; I have no idea if these picks are still fast, no promises they are still the same price (pretty terrifying honestly).
I'm mostly on coding plans so it doesn't super affect me. But man is it a bother to maintain.
The point of this loss leading is to properly hoover up the money in the pockets of enterprise customers, get them locked into the idea that they need the latest and greatest cloud-based model, while simultaneously starving everyone of the memory they'd need in order to run competent models locally.
In not-too-distant future we're going to be running better models on our phones than we can buy access to today in the cloud. Skate where the puck is going: soak the customers until that day comes.
It's interesting that the cost multiplier for Claude Sonnet 4/4.5/4.6 varies so much (1/6/9), while the API cost is exactly the same for all three models.
Also, the multiplier of 27 for Claude Opus 4.6/4. is way higher than the increase in API price would suggest.
On GitHub copilot you pay per prompt. More powerful models can do a lot more work (consuming a lot more tokens) per prompt. Also, they tend to use more thinking tokens.
> More powerful models can do a lot more work (consuming a lot more tokens) per prompt.
That is not my experience. Each model since at least GPT-4 can fill up an entire context window.
In fact, more powerful models can solve tasks faster, so their ratio of multiplier to API price should decrease, not increase.
For example, Claude Sonnet 4.6 has a multiplier of 9 and an API price of $15, which is 0.6 multiplier per dollar.
Claude Opus 4.7 has an API price of $25, so it should have a multiplier of 25 * 0.6 = 15 when extrapolating from Sonnet, but the multiplier is 27.
> Also, they tend to use more thinking tokens.
That might be it. Is there any data on this somewhere?
Those multiplier are only for grandfathered Pro an Pro+ plans that had annual billing, basically a way to scare people of out of those plans.
Ant new ones (and bussiness+enterprise plans) will be on token based billing since June 1.
One theory of the play of SpaceX might do if everyone migrates to query-based billing:
Provide cheap and unlimited access to Grok for programmers (hence the Cursor partnership/purchase for distribution).
-> This would drag massive revenue right before the IPO announcement, like if the company is super growing
-> At a loss, but don't worry, we need these funds to build the biggest datacenter of the universe.
This announcement would create enough momentum to increase valuation, and because of the merge of his companies, would save his X/Twitter investors from a tragedy.
-> Would also be a great service to Cursor investors and so, who are stuck with their VSCode fork
Which in turn owns Twitter. SpaceX is now a social media company in addition to a rocket company.
One theory I think Matt Levine posited, is that SpaceX will go public with dual-class stock that gives Elon control even with a minority ownership stake, and will subsequently buy Tesla, which doesn't have dual class stock, making SpaceX the singular "Elon Musk company", with him having operational control despite being public.
Show HN timing matters more than people think. Monday-Thursday, 9-11am Pacific, is when the front page has the most engaged readers. Weekend posts get less competition but also less engagement.
The routing automatically routes you to other inference providers (for the same model) if/when the original provider goes down.
It's a convenience cost, for sure, but it's not valueless in a fast-moving world. Certainly if you're comfortable with one provider and it's cheaper, do that.
For me the largest value-add is the unified API. Being able to instantly start trialling a new model with zero code changes is well worth 5%. The other part is not having to deal with billing for multiple platforms.
What's annoying is that it's obvious. In the case of GPT 5.5, if Copilot is going to charge 7.5x what GPT 5.4 costs while OpenAI themselves via the API/Codex only charges 2x of what GPT 5.4 costs, that will immediately raise an eyebrow.
To anybody who's been watching the tech sector with a critical eye for pretty much any period from the late 90s and onward, this is just the enshittification process. For most of OpenAI's existence it's been obvious, to me, that investors were burning insane levels of capital to build the market, and now that folks are locked in, you're seeing higher fees, ads, etc. Yet again, the user is the product; the investors want to siphon your data, attention and once you're hooked, money. And for companies like Microsoft and Apple, those hooks can dig deep.
Let's call it for what it is dumping. Dumping things on market below cost of production. This should not have ever been allowed. RnD costs I can accept somehow. But in this case the interference should have always been billed for the real costs that it took to produce and pay off the capex.
If you paid attention to the power requirements and amount of hardware being put into data centers, you should have realized that it cost them an order of magnitude more than you were being charged. To rework your analogy: they hooked you, now they're gonna see if they can reel you in.
They can only reel you in if its worth it. I still can code.
And while i do not spend 200$ privat, in my startup we discussed this and our current mental model is, that instead of hiring someone new, we prefer to have more money for tokens.
This is easier for us and has a bigger benefit. The cost of a new / first employee is very high, a 200$ subscription is not. Upgrading that to lets say 400 or 800$ is still alot easier and if i can run multiply and better agents with that money, lets goooo.
I'm looking at education -- teachers and students, not terribly tech savvy, are being mandated to use these tools. And then comes the rug-pull. It was worth it, but now it's outside of their budget. Poorer schools / students can't stay at the cutting edge; richer schools / students can.
“Enshitification” is just when unsustainable subsidies end?
Another reason to hate that word.
From a different perspective, you were granted an incredible gift from the companies who let you use their product on their dime. Hopefully you made the most of it when you had the opportunity.
No, it's much more than that. It starts with unsustainable subsidies, as Uber undermined the taxi industry with a ludicrous burn rate. And then, once everybody's hooked to the point that they can't imagine life without the product, you raise costs. And you iterate: raising costs, lowering quality, selling data, increasing addictiveness. Until everybody wants to get rid of it, hates every aspect of it, but is still hooked to the core product. I'm personally not using these tools, not using uber or Meta products. But I'm still using some Google products and it's hard to extricate them from my life now that I'm using them.
OpenRouter doesn't even have hardware. What are they possibly subsidizing? The platform costs?
OpenRouter is guaranteed to be about the highest margin operator in the business right now. Everyone wishes they'd be them, skimming 5% off as the middleman without any OpEx.
> OpenRouter is guaranteed to be about the highest margin operator in the business right now. Everyone wishes they'd be them, skimming 5% off as the middleman without any OpEx.
The 5% fee probably has to factor in Stripe's fees, which would be around 3% to 4% depending on whether it's an international card.
Streaming, caching, and tool calling can get pretty expensive with scale, even when you don't touch inference. Maybe they're doing something clever and are quite profitable.. or maybe they've already taken $40mm from VCs and are currently trying to raise $120mm at a 1.3B evaluation.
They also show headline prices for the cheapest provider of whatever model, but then need to hit different backends some of which may be more expensive. For now they absorb those costs, but the VCs always come knocking.
Just my opinion though. Totally agreed that they have one of the best positions amongst all AI providers from a financial standpoint.
> They also show headline prices for the cheapest provider of whatever model, but then need to hit different backends some of which may be more expensive. For now they absorb those costs, [..]
They do?? I was under the impression I was just playing the price for whatever provider they deemed 'best' for each completion.
Checking now: The way they describe it in their FAQ is that if the price changes, then they will bill you the new price. But I read that as regarding if the primary model provider changes their headline token cost; not in the case of pricing differences for models that have many different backends that host them.
Regardless, I would be more concerned about the streaming costs if the service continues to blow up and they scale aggressively through VC investments. If their 5.5% skim accounted for what they needed, you'd think they could effectively grow organically..
I am and I see it as stopping the music at a party when you want everyone to go home without telling them to go home. There is also the offer to quit with prorated refund for the remaining time. I think I am going to take it.
I don't know if it's just me but copilot kind of sucks. I've been running local models with like 9b parameters and they are about as good if not better. Obviously there's no integrations or whatever and I get most people are probably paying for that than anything else but eh. Big no thanks from me.
That's so unfair to us hard working developers. A month ago i could buy for .4$ a turn with Sonnet. Now i have to pay at least .9$ for this turn. Weeks ago i could buy for .12$ an Opus turn after they already raised prices and now they want .27$ from me for the same product! They are stealing from us!
They aren't stealing from us, for several reasons. First of all, it's a voluntary transaction. If you don't like the prices, use something else. Or don't use AI at all.
Second, you have no idea what their costs are. It is most likely that they are simply passing on their costs to you. If that was not the setup, users would just go to another service provider who was providing tokens at a cheaper rate. It's not like there is a dearth of competitors in this business.
"Your plan pricing is unchanged: Copilot Pro remains $10/month and Pro+ remains $39/month, and each includes $10 and $39 in monthly AI Credits, respectively."
If there's no discount on credits (in terms of tokens per dollar) over other providers, I'm going to switch to a PAYG provider. If there's a month where there's little to no coding I can pocket the 10$. What incentive do they give to stay with this plan?
Or if you're a business with multiple seats, these plans may be more inefficient than raw API usage billing. Since if anyone at your organization fails to utilize their full $19/39 allotment each month, that's wasting money, whereas with API credits it is 100% utilized.
I don't think they've thought through the implications of this. Everyone should cancel and go usage-based billing with caps.
They do address this in the doc, Orgs can now (although it was vague as to whether it was an option or just the new standard, probably option due to business contracts) 'pool' the Usage billing across all users.
I'm guessing they did that (and the 'temporary bonus credits') to make the pill easier to swallow for that side of customers.
It forces you to pay at least $20 in tokens per user even for people who use less (they probably have stats on how many people use just autocomplete, which doesn’t count against the quota. or have a seat and don’t use the service at all).
You can get at least at baseline vague stats but from what I have seen it is more account than user focused, i.e. 'X number of users used %feature%' and/or '%model% was X percent of requests per day'.
That said it's worth noting, I don't see how anything they expose will reliably help orgs plan costing from what AFAIK is in fact a big shift for billing/costing planning.
> It forces you to pay at least $20
For better or worse that's public pricing, i.e. if you are coming in also negotiating VS for devs, windows/office licenses for the rest of the business and stuff like Azure Devops... a lot of their stuff gets cheaper if your company's IT procurement group is vaguely competent at negotiating. Not even talking bigcorp here I'm talking 500-1000 employee range.
Of course, very small orgs will suffer, but it does tie in with the theme over the last two weeks; anyone with a personal account is basically subsidizing the credits for the business accounts during the transition period.
This was my first thought too. "Oh cool, I should be seeing lower prices" as I don't use Co-pilot that often anymore. But no, that's not the case. It rather served to remind me that I should probably just cancel.
Everybody who says it's a 5-9-27x seems to not be aware of the obvious loophole. More like 50x increase. You were able to use over $500 worth of Opus on a $10/mo Github plan easily, no hacks. You could just prompt "plan this out for me, don't stop until fully planned, don't ask any questions", and you would get ~$5 worth of planning in one 3x request. At 100 requests/mo, each easily reaching $5, that's easy $500 worth of tokens.
It was going to happen regardless due to the nature of enshittification. If they really wanted to stop people using 100M tokens a day, they could've prevented it years ago.
So, silicon valley decides to use their playbook of expand at all costs by burning money to acquire the market (like a carcinoma), and it is the users fault ?
Should we be blamed about uber destroying the taxi business, or airbnb the hotel one? Oh sorry, "disrupting".
Uber was dirt cheap, now it is the same price as taxis, and the people working for it (the "partners", not employees) have no social benefits.
Airbnb was cheap and humane, now it is THE cause for housing crises and massive residential property "investment".
The playbook of silicon valley is destructive, not disruptive.
It is by design aimed towards wealth accumulation. The ones with most money can capture the market, and make even more. It really is late stage capitalism.
And the more wealth inequality there is, the more pain, poverty and instability will be as well. AI will only exacerbate this.
Yeah it was crazy. Nowadays I use pi with OpenAI GPT 5.4/5.5, which to me seems both better and more generous than Claude. I supplement it with OpenCode Zen to get access to a bunch of models at token cost, and OpenCode Go ($10/mo) to get subscription-style access to Kimi, GLM and friends.
That was not my experience. When I tried to use Opus for longer tasks with Copilot, it would fill up the context completely and then crash without any output, while still consuming premium requests. (At least from September 2025 to January this year. Haven't tried after that.)
Copilot has improved immensely in 2026. I'd say to give it a try again if you're up for it. It works about as well as Claude Code these days in my experience.
Unfortunately, Opus was removed from the student plan in March. So far, I had been happy with GPT-5.3-Codex, but that model seems to have been removed this morning.
When using opencode or copilot CLI, the error messages are displayed
normally and it's possible to see what's going on. Under Pi, it
sometimes just hangs, or Pi crashes with some bun stacktrace and that's
it.
Copilot has introduced additional limits for Claude models in past
month, and it's rather easy to hit it. Pi often doesn't show anything
when this limit hits (although sometimes it shows the error, I guess it
depends on Pi version).
This was my solution to very very though compiler tests that would take sometime up to 4 hours to figure out. Some of the time would be spent on running the tests, but still... I was burning so much tokens. I have free Copilot for my open source work so I wasn't even paying the $20.
The 27x multiplier on Opus is the tell. That is not a pricing model designed for broad adoption, it's a price signal that says 'use the cheaper model.' The problem is that once users start self-censoring which model they reach for based on cost anxiety, you've degraded the product experience in a way that's invisible in the metrics but very visible in churn.
Flat subscriptions had one big advantage: zero cognitive overhead per request. That's worth more than people admit.
This may be a more accurate analogy... "The Porsche you rented at $200/mo now only allows you a maximum of 100km of travel. You will be automatically charged extra when you go over that."
On top of being worth less, the subscriber discounts are gone.
The old plans were $0.033/request for Pro, $0.026/request for Pro+ and $0.04/request for pay-as-you-go. That discount is now gone. They even still advertise "5x the number of requests" for Pro+ over Pro.
Yeah, if I go to a petrol station with 50€, but only get a tenth of the amount of petrol I got last week, I may think that the price has in fact changed.
It’s technically true that the plan prices haven’t changed, it’s just the value you get from those plans has plummeted. It’s classic deceptive sales language.
I was curious why a company would still use the VS Code + Copilot sidebar method for coding, rather than something like Claude Code. Turns out there’s a GitHub Copilot CLI!
I thought I was pretty familiar with available options, but no one in my circles ever mentions this product. It doesn’t seem to have much mindshare.
I'm curious about the opposite: Why would anyone use the CLI when, at least with Copilot, the VSCode plugin is super tightly integrated with VSCode, meaning the agent can see everything I can see. There's no mismatch in linter calls where I can see a lint in the ide that the agent can't find for example. I've had this problem even using CC in their VSCode extension, so I can't imagine it's not an issue in the CLI as well.
I use the Claude Code VSCode plugin for 80% of my work.
I prefer it because I can look at the code (although not as often anymore) and config (very often!) easily.
It also lets me jump to previous conversations easily.
There are a few cases where the CLI makes sense. One big one is if you are running multiple simultaneous sessions on a remote server using Tmux to have them preconfigured when you reconnect is nice.
The vs code integration is pretty slick. I can copy and paste function names into the prompt and it automatically turns them into these `#sym:` reference objects that I presume populate the context window with metadata about the function and where it lives. It knows what file I'm currently looking at as I jump around in the code, and that automatically gets loaded into the context. I can also drag and drop folders or specific files for context into the sidebar.
It's a lot of stuff that makes me have to type less into the prompt, since it's already getting so much info from my editor
I’m actually trying to move back from the Claude Code style, I feel like it’s easy to become distant from your own code, and I am feeling uncomfortable with that.
I’ve “vibe-coded” some projects and when I start to find issues or go to refactor them I don’t have that memory of why decisions were made, because many decisions were never made.
I've used it quite a bit. There are a lot of AI terminal coding products and this is another one. It works well, handles sub-agents without issue and does a reasonable job operating in the Copilot ecosystem. It handles mid-task questions and such we well.
But its a really good UI for agentic coding. Not sure why more people don't use it. I've tried the others and keep coming back to Copilot chat. It's a really good tool. Which is why the rugpull on pricing is so concerning.
Yeah, I've been using it heavily at work since the beginning of January (and have a personal Anthropic sub to compare to). Copilot CLI is pretty good, honestly. Most new features in Claude Code get cloned by Copilot CLI within a couple weeks. Claude models seem mildly more clumsy in that harness than the one they're trained on - subjective guess around 20% more turns for an equivalent task - but it's not a noticeable difference in the final output.
The other cool thing is Copilot SDK, so you can build agentic capabilities into apps, or build tools, that leverage the agent harness of the Copilot CLI:
> I was curious why a company would still use the VS Code + Copilot sidebar method for coding, rather than something like Claude Code.
I use Claude Code, but I kept my Copilot subscription around mostly for really cheap usage of other models when I need to try a different one (which appears to be ending, in a sense) and also the autocomplete in Visual Studio Code which was really great across a bunch of files, I could make changes in one file and then just tab through some others.
I wonder what other good autocomplete is out there.
> also the autocomplete in Visual Studio Code which was really great across a bunch of files <...> I wonder what other good autocomplete is out there.
I am in the same boat. I tried looking for tab/auto-complete implementations ~ a year ago and it was pretty disappointing. If that has changed, would love to know!
Windsurf has free unlimited tab complete, I use it as an IDE alongside Claude Code and it works pretty well. I think Google's Antigravity also has free tab complete but no idea if its any good.
That's espoused as the big reason for the price increase: most Copilot subscribed developers it seems have moved to "agentic usage" with the CLI and Cloud-based agents.
Which feels a bit like a kick in the pants for me as a developer that was primarily using Copilot for VS Code ghost text and very rarely used the Chat sidebar much less "agentic" tools.
Copilot Pro sort of made sense for my personal account when amortized across a year, but I don't want to "waste" $10/month on credits I won't use most months.
Liability. I work in a tightly regulated market and I cannot use any service due to be in touch with PII data. If it was my option I would just use OpenCode with deepseek, it is more than enough for the majority of my tasks.
I'm just so confused why people aren't just using ghostty/kitty/terminal.app and claude code. Compared to the other approaches I've tried, it's by far the most effective way to get performance from opus 4.6/4.7
I don't know about others, but I use Copilot more often than other apps because of its tight integration with the VS Code itself where I still spend most of my time working on other things while letting AI do some task that I decided to delegate to it.
I tried the VS Code + Copilot sidebar approach a few months ago. It was definitely rough around the edges compared to Cursor/Claude. In our corporate environment, we weren't even able to use frontier models.
I've been using it for a few months because Copilot was the only AI blessed by our corporate overlords. It's not bad, I would say it's about 80% as capable as Claude Code, which I've used extensively on personal projects. However CC was recently approved, and I'm betting that with these changes to Copilot pricing we'll end up dropping it like a hot potato.
Search has become so bad that I also struggled to find Claude Code alternative and made my own tight (not editors, not plugins, not agents, strictly similar to Claude Code CLI) list: https://github.com/omarabid/cli-llm-coding
The list is not long but there are quite a few options. Even Grok has its own CLI!
The reality is, even though a CLI prompt looks very simple, it's a very complex piece of software. I personally use Claude Code (with GLM) and anything else I have tried was significantly inferior (with the exception of opencode).
I don't use Copilot or any paid AI but all of this usage-based billing reminds me of cellphones back when you paid per individual text message.
Usage paying for AI is 1000x crazier because you're not even getting a guarantee in the thing you pay for in the end. You have to keep feeding it prompts and hope it gives you the solution you want. You may end up with no expected result yet you are paying for it. At least with texting, you got what you paid for.
I wonder how long it'll be before all AI costs are flat unlimited monthly fees or even free across the board, without compromise.
I expect in the future we'll find out that someone in the industry was juicing the numbers with fake thinking tokens or something. The whole pricing model of charging you for the tokens it generates while not knowing how much it is going to generate going in has always been pretty crazy.
It reminds of early smart phones when the cell providers pulled away from unlimited data...and then they brought it back in s few years.
I think competition will get fierce. We see many people are attracted to the price stability of GHCP - it became clear what a request could do - the problem is that they didn't match results with cost. It's not clear what a 5 hour usage window in Claude Code can do.
There's no reason the harness couldn't provide a quote on the next request, aside from it takes effort and it would be upfront to the user, creating expectations.
Yeah, this was my frustration with Suno and Sora. You can burn a lot of credits (not to mention time) generating things that aren't what you wanted.
I don't mind a PAYG model for a simple chat interface. But when it comes to actually producing things, you burn through TONS of tokens creating the wrong output.
It incentivizes you to do most of that prompting on your own hardware/time, and only feed the final prompt with only necessary context to the big AI in the sky. It might even force you to think about the problems yourself for a bit!
> I wonder how long it'll be before all AI costs are flat unlimited monthly fees or even free across the board, without compromise.
That's already the case if you can self-host an LLM; you don't even need a mythical H200: gamer-grade GeForce cards can get you a long way there (if this page is to be believed: https://www.runpod.io/gpu-compare/rtx-5090-vs-h200 )
...after RAM prices return to normalcy, of course - and then wait another 2 or 3 generations of GPU development for a 96GB HBM card to hit the streets - and also assuming SotA or cloud-only LLMs don't experience lifestyle-inflation, but I assume they must, because OpenAI/Anthropic/Etc's business-model depends on people paying them to access them, so it's in their interests to make it as difficult as possible to run them locally.
That page compares models that easily fit inside the ram on either GPU. The biggest difference comes when one card can fit a model and the other cannot.
> In March 2026, Windsurf replaced the credit-based system with a quota-based usage system. Instead of buying and spending credits, your plan now includes a daily and weekly usage allowance that refreshes automatically.
With hindsight, per-request pricing makes no sense at all if an agent can burn a widely varying amount of tokens satisfying that request. These pricing plans were designed before coding agents changed the dynamics of token usage.
I wouldn't call it hindsight - I don't think anyone, at any stage, thought running a 10 minute+ sonnet session for 1 premium credit was ever profitable. We all knew it was a loss leader to get people using it.
It would have been profitable if that premium credit cost more than a negotiated discounted rate with Anthropic. We have no way of knowing if there were negotiated rates though!
There is no way to make that cost model profitable consistently. If 1 prompt can mean 100's/1000's of requests over hours, and you only pay for that 1 premium prompt, that can never be profitable.
They can engineer the harness to limit the amount it does. When pressing enter, it's be nice to have a "budget" per prompt, much like the model multiplier. When the harness used up the budget, it cleans up and cuts off the work.
But that would entail actual work and effort...and care for user's time and money.
Guys, you're discussing a house of cards to begin with: No matter how you're paying for the $CURRENTSOTA you're not garunteed that next month what you pay for will be the same.
So, lets do some honest evaluations:
1. The model itself is a non-deterministic engine of work with an unknown value; it's real value is just magic.
2. The business model itself is non-deterministic engine of profit with a known value; whatever the VCs have put into it, _must_ be piulled out. If Ed Zitron's numbers are correct, circa 2030, it's several trillion dollars.
So do some matrix multiplication of non-determinism vs determinism, and realize that the value proposition for _you_ is only going to decrease because #1 can never outpace #2, ensuring enshittification captures a smaller and smaller whale.
We know this. This has been the last 2 decades of money extraction from software. It was ok when it was some 12 year old's parents CC. But now it's you, or your business, that's going to either ben squeeze for value or squeeze out of the market.
And everyones squabbling about the color of the cost.
ok
The problem with assuming that tokens can only get more expensive is that the Chinese open weight LLM firms have dropped models which have a known, fixed price that can never get more expensive (since we can run them on hardware we own).
Well, I guess we're not discussing the same thing. The cost of cloud tokens are going to go up. They won't ever be cheaper. They're generating far more tokens than my AMD 395+ w/128GB at a much cheaper rate.
I agree though, it can't get cheaper than the cost of hardware it's just without sufficient documentation of the actual costs to run the cloud models, we can't really know what the "true" cost of each token is. I assume there's an economist out there somewhere that could figure it out though. Certainly, the cost should approach at a minimum a open weights model running on a local machine.
I've succesffully got Qwen3-coder-next to loop and generate sufficiently competent code and from what I can tell, the difference between this and the cloth is how quickly the gen happens and perhas how interactive it has to be.
I wonder if GitHub (Microsoft) is implicitly betting that enterprise demand is sticky enough to absorb these rates, especially given that Opus 4.6 “fast” was being listed at a 27x multiplier. Maybe they saw enough usage at that price point to conclude the demand is real. Or maybe the strategy is to keep the enterprise customers who can justify it while shedding heavier individual and power-user usage.
The interesting question is how long it takes enterprises to notice the capability/pricing tradeoff, and whether they respond by limiting access to the strongest models internally.
The part that worries me is that this market is still very early. Most developers and organizations are still learning how to use these tools effectively. Raising the experimentation cost this much may slow down the discovery process that makes the tools valuable in the first place.
As someone who works in a Microsoft shop with Copilot, I am curious to see what happens.
Due to data governance it will be difficult to move to a different provider.
At the same time, this price hike is so large that the ROI on copilot will be a net negative.
I think what will ultimately happen is that we will not pay Microsoft more than we currently do and we'll simply end up with less AI usage in the company and a reduction in productivity.
I also suspect that there are many "slow-moving", Microsoft heavy enterprises but with in-house devs that can't get anything but Copilot approved, and Microsoft trusts this will remain so.
It's not turning consumption based because there are a ton of these licenses just sitting idle.
As a single data point, this is absolutely true. At my current "Big Corp", Copilot was immediately approved while Claude is entering month 2 or 3 of trying to get approval.
Additionally, we got copilot for every user, including those that never write code or use AI tools.
yes. even in my province that's relatively behind on tech, at least +-20% were already strongly relying on Microsoft (Azure) and thought adopting Copilot was a no-brainer. That ofcourse comes with autonomous programming, because we're paying for it anyway!
These companies already lost the spark/innovation years ago. They're just using LLM's as a way to survive. Wonder how long that lasts.
As someone that is on the enterprise side in a non-tech F500 company, what I'm seeing is some FOMO and need to be part of the hype cycle. We're about to plonk a bunch of money on more Copilot licenses. Something got in the water where all the C-levels the past two months are pushing everyone to use AI but when they bring up examples of their uses its like "I use it to rewrite my emails" or prompt 'engineering' ideas that point more to patching over poor processes, data management, and decision-making within the organization or not.
What we're seeing across the board is every software company tossing AI onto their name or sales pitch and no one understanding what that actually means. But we will spend money on it because of FOMO.
I really question if we're reaching the end of the hype cycle to the point. I wish I were brave enough to put money on it. It feels like there was a command from up top to 'do something with AI' and leadership is scambling for some resume-building projects vs doing the hard work they should've done the past two years at a people and process level.
I'm at an org, where every medium to large company meeting talks about AI in similar breathless non-specific terms, yet m365 copilot is the only approved AI chat, and GH copilot the only AI coding tool.
Subsidies stop when LLMs improvement plateaued (though they still benchmark higher somehow). At some point, you have to make money or at least break even; and I think they concluded that we reached that point.
I really hope that they think so and that they're wrong and they get burned hard. Them and all the AI labs that lied, stole, inflated, hoarded and tried to justify all this as an existential moment where AGI would radically change society. I hope their calculus to reel in paying users is all wrong and now they all crash and burn instead of recouping VC money.
Github had, by far, the most easily game-able agent usage policy. People would force the agent to run a script before the end of turns that consisted entirely of `input("prompt: ")` so that you could essentially talk endlessly to an agent for the price of a turn. I see this less about the future of this industry and more about fighting the costs incurred by bad actors.
I never played any games like that, but simply giving the agent a clear exit criteria and instructions to check the exit criteria every time it thinks it's done on a complex task was often enough to keep it chugging away for most of a day on a single prompt in my experience. Per-prompt pricing just isn't sustainable period, even if everyone is acting in good faith.
I once asked it to do a comprehensive security review of our code. It churned for nearly an hour (and then produced 90% false positives). Insane that that usage was charged the same amount as me just saying "Hello".
After 2 months of using copilot pro, they've charged us 22$ in premium requests for my user, i spend roughly 1900$ in tokens, going of the on-demand pricing. This is estimated to be around 40-60% of real costs. They are undercharging by a factor 50 to 90!
One reason I used it was that I wasn't locked into a single provider and switching them was as easy as changing a drop-down. Small feature? Sonnet or GPT5.4/mini? Large changes? Opus. And why not see how good Raptor Mini does this one refactor?
It also helped build an intuition of what wach model could do and which parts it was weaker at because you could try them almost side by side, especially if one model's output wasn't great.
That said, these were all side projects so nothing truly consequential. Otoh, you might leave some extra perf on the table but I found the models worked quite with the Copilot harness.
Yeah, this is a very useful abstraction layer. The entire concept of separating the model creator from the model runner is good for competition and is customer friendly. Which means they likely hate the concept and want to kill it.
Gosh, imagine getting to do that with your TV/Streaming subscription. Getting to pay one fee to access some set number of hours per month from any of the providers.
The problem is I can't afford the tokens! Even on my $10/mo plan, running either 100 opus, or 300 sonnet agent runs would cost hundreds of dollars - well above my budget!
I liked copilot because I didn't have to think about tokens. I get hung up when having to think about the price of things, and its hard to think about the project at the same time I got to think about token usage like a gas bill. The usage system had its own issues, but having a set amount of requests was a very comfortable way to use a paid AI service.
Not paying per token? Not sending my code to someone else's servers for inference? That's the stuff of sweet dreams for a stingy, paranoid solopreneur like me.
If I could run a local model comparable to even Sonnet 4.6 without shelling out $50K in hardware, I'd do it in a heartbeat. But all I have is a 32 GB of RAM and an old RTX 4080.
Or am I not up to speed? Are there decent coding models that can run on dev laptops? Not that that's what you were suggesting by recommending a local model, necessarily; just curious.
I am trying to figure this out too... what I am seeing is that the local models like Qwen 3.5 family that fit on hardware like yours handle ambiguity poorly. But are capable of emitting complete apps too.
I do love using local models when I can, but qwen-35B is the best model I can run, and while its an insanely good local model, it does not compare to the big ones.
So given that I primarily interact with LLM's through VSCode, and I prefer the Copilot interface to the Claude Code plugin, does anyone have any suggestions on other plugins I should try? In my experience, Copilot is much more "plugged in" than any of the other plugins, in the sense that it can see things like linter outputs in VSCode. Basically, copilot "sees what I see" in a way that no other plugin or command line tool can, which make it much more ergonomic to use.
With this pricing change, I see no reason at all to stick with Copilot in principle, but I really need to solve this issue of IDE integration to move on.
You can use Copilot Chat* with basically any API provider, and if you switch to the VS Code Insiders build you can configure it to use literally any OpenAI API-compatible endpoint.
Other than that Zed has a similar experience which is pretty decent.
* By which I mean the good one, whatever it's called now - the part of Copilot that used to be a plugin and is now part of VS Code, not the thing that has always been part of VS Code.
Simply the output of the code written. It initially worried me that it was harder to add context (can't select a single line of code in VS Code and automatically attach it, for example), but I've found that the Claude CLI's output is so far superior that it doesn't even need that VS Code context.
Cancelling. Going with Codex $100, Kimi annual plan, DeepSeek API, and a local LLM once I get a Mac Studio.
Inference economics are going to be brutal in 2026 H2 when DeepSeek's new infra and model improvements come online, and Kimi launches K3. By brutal, I mean for OpenAI and Anthropic.
Yes there's no Opus at all on Pro. GPT 5.5 is also missing. Then again what would you expect, the economic reality is beginning to hit. Also I can't be too mad when the "base" models (GPT 5.4...) are still available and decent.
When I see how fast Codex max thinking GPT 5.5 eats our enterprise seat credits almost anything else seems cheap (until we switch our live systems from 5.4 api to 5.5 api I guess)... good thing I'm not the one paying for those credits and tokens (which is probably how most of the money is going to be made on AI going forward, borderline free chatbots for normies are done)
On my personal account, Copilot Pro+ still only gave me back Opus 4.7, whereas my work's Pro account still lets me use Opus 4.6.
So, my gut says, it's entirely possible that Pro+ will continue to have more segregation on model availability...
FTA
> Last week, we also rolled out temporary changes to Copilot Individual plans, including Free, Pro, Pro+, and Student, and paused self-serve Copilot Business plan purchases. These were reliability and performance measures as we prepare for the broader transition to usage-based billing. We will loosen usage limits once usage-based billing is in effect.
There's enough weasel wording here that I would expect only certain models get re-enabled on Pro.
e.x. lots of people seem to get good enough results from Opus 4.6, personally I prefer it over 4.7 in GH Copilot... locking that down to Pro+ would be, given this salvo of enshittification, a 'logical' move on their part.
You should indeed feel vindicated. Me too, I've been a paid subscriber for months and eagerly read every word. <wink> But yes, I feel like this is the first of several vindications you have coming.
What's the current situation for coding with Local LLM's on decent hardware? I have an M3 Max with 64 gb of ram and am thinking I should start looking at Ollama and Opencode? Is this a useful stack for smaller personal projects?
It’s getting there. You could give a try with qwen 3.6. It’s worth paying for better models in the cloud, but local models are now better than nothing.
One nice development recently was ollama's support for MLX optimization on Mac hardware. It's not obvious how to know you're using a model that works with it, yet, so it's rough around the edges.
I think that only applies to held-over users on the annual plan:
> Users on annual Pro or Pro+ plans will remain on their existing plan with premium request-based pricing until their plan expires, however, model multipliers will increase on June 1 (see table).
It isn't just the big multiplier increase, they also say "...and no new models or features will be added to annual plans going forward."
Can you imagine ten months from now and you're still rolling Sonnet 4.6?
Cancel/refund is looking pretty good. They're doing refunds until May 20.
"To request a refund, go to Settings → Billing and licensing → Licensing, select Manage subscription, then choose Cancel and refund "subscription". (The phrasing varies slightly depending on your subscription ). This option will be available until May 20."
It is an apples for apples comparison since those new multipliers only count if you are on an annual plan in which case the premium request system stays in place until you either cancel and get a refund or until your renewal comes up. https://docs.github.com/en/copilot/concepts/billing/usage-ba...
Those multipliers will only apply if you are currently on an annual subscription (and only until your renewal comes up or you cancel). So I assume they simply want to make it as unattractive as possible to get most people to cancel it and move to the token based system.
That's not an answer. It's specifically a discrepancy between 5.4 and 5.4-mini. If you look at all other models/generations you see that the cheaper model indeed has a lower multiplier. It's very strange that only 5.4 doesn't have this.
They're not the only ones in the AI sphere to wind back, but they're the weirdest case in my eyes. Microsoft invests in having engineers building open models and they don't use a single one. I really don't get it.
But what really surprised me most about Copilot is that it would bill you per question, nothing about tokens. So if I managed to produce a prompt that gave me back an insane amount of tokens for something, which using any Claude model would easily accomplish, you were giving me my money's worth, at your own expense. The math is not gonna math out forever.
One of the largest employers publicly engaging in a project which has the outcome of depressing wages. It's easier to "get" if you don't take the trillion dollar gorilla at face value.
The cheapest copilot plan felt totally unsustainable to me. For around £8 month i was getting 100 opus 4.6 prompts (albeit with a reduced context window size around 128k iirc vs 200k to 1m for first party hosted opus). Gpt5.4 was hosted with 400k context iirc.
On top of that, you’ve got 2000minutes of container runtime, so running cloud agents was included. As was anthropic agent sdk mode via copilot which is very comparable with claude code - not identical, the anthropic “modular prompt” is much leaner in the sdk version.
I cant say im mad, i got above what i paid in value. That said, going forward ill probably go back to openrouter payg rather than a subscription.
I got a free 3months of the gemini £19 plan and ive been playing quite a bit, 3.1 pro is a good model, i just find it slow. Flash i think i under appreciated until now.
I AM mad, because I just signed up for my private copilot sub. Otoh I do have it at work. I suspect we will start noticing differences there, too, but it all depends on the top users, vs average user since it is now pooled, IF they dont just allow extra usage. Bu that has become more expensive now, so not sure what they will do. I am hopeful that very few actually even use it. In my and surrounding teams, only maybe 1 out of 3 people really use it much.
I pay for Copilot annually, and mostly for its code auto completion features. I use CC if I want to do anything agentic. Not sure if I want to pay more for occasionally-good-intellisense at this point.
But you can no longer amortize annually, which makes it even more a question of "is this worth it this month?" each month. Especially for personal accounts.
I'm similarly thinking about sticking with the auto-downgrade back to Copilot Free when the annual sub ends and then just yelling about it any months I hit the 2000 completion cap.
I wouldn't mind a plan between Free and Pro that is just "all I care about is code completion and next edit suggestions".
I am seriously considering binge buying local AI inference hardware. The way this is going, there will be another big GPU crunch soon because everyone will need local models and/or open model inference capacity to do their programming tasks when the subsidized subscriptions are no longer flowing.
I doubt you can force them to provide the service with the original terms, but you might be able to ask for a (partial) refund. If not today, after a week of verbal abuse they will receive for this online.
It depends where you’re located. In the EU they have to honor the contract you entered, but presumably there is a clause that they can prematurely terminate the contract without cause and give you all of your money back (from the start of the contract).
In places with reasonable consumer protections (Australia, Germany) it almost certainly is illegal unless they give a full (whole year) refund. I think the short time limit of applying for a refund won't be looked at favorably either. Regardless of their ToS which I'm sure covers this.
But companies do lots of illegal things, and in general nobody takes them to court over it.
My thought exactly! First the usage limits + model limitations and now fundamental change to the billing. Hope some consumer watchdogs are looking into this!
That kind of clause would be void in many places around the world.
For example, the German Civil Code states:
Section 308 - Prohibited clauses with the possibility of valuation
In standard business terms, the following in particular are ineffective:
[...]
4. (Reservation of the right to modify) the agreement of a right of the user [TL note: this means beneficiary of the terms, eg. party or other subject of the contract] to modify the performance promised or deviate from it, unless the agreement of the modification or deviation reasonably can be expected of the other party to the contract when the interests of the user are taken into account;
"It" being the end of subsidization of tokens and plans (expected) but while lock-in to foundational models and cloud services is still lacking. Guess investors want their ROI sooner than later, given how big of a wrench the AI boom has thrown into global economics.
It's almost impossible to afford the usage-based pricing as a heavy user with multiple max and pro subscriptions. Seems like a bet that same-capability inference will keep getting cheaper on something like a Moore's law curve.
“Base plan pricing isn’t changing” is technically true, but for anyone using the more capable models heavily, this is still a price increase in all the ways that matter. The old abstraction was hiding compute costs; the new one mostly stops pretending.
Anyone paying enterprise billing can you shed some light how on the earth you will be able to fund this bill going forward?
Were you able to see assisted AI coding savings proportional to costs increase now you are going to get?
Companies removed people as AI assisted coding will be cheaper and now coding cost are going up from fixed $X to non-deterministic. The posts by Uber few days back about spending 12 months' worth of money in 4 months tells a lot.
Only path forward seems using Open-source models and many companies don't use Chinese that makes only Mistral one as the option.
Surely one path is to return to humans. We can accept that a profit grab to reduce staff numbers to maximise profits, maybe was never the best plan without being in control of the development tool.
Sample of 1 opinion: Massive downgrade for fun vibecoding on the go. Got done much before the recent rate limits on mobile on fun projects, their product went from fun to bad within a few weeks. And now into oblivion I suppose.
That was the first thing I turned off in VSCode. Autocomplete for my TypeScript projects was great. And the "AI" suggestions/completions were really getting in the way of me still being the "driver."
some of Github's open source maintainers have lost their free github copilot pro, guess this is really the next step for them to save cost in their infrastructure.
I started to use github copilot with vscode, but have never been too happy about the system. Over the months I gravitated to much more agentic workstyle, hardly ever editing much code by hand. The vscode IDE was getting more in the way. I had already started to look at OpenCode, and when I found it has a web interface, I was happy to switch over. I use a simple editor (KDE's Kate), or just less to skim through the code and/or a git diff.
OpenCode has some free models in it, but I think I will need to get some kind of subscription for a better one. But it won't be copilot any more. The market is moving so fast that I don't know what are the most resonable models, or the most flexible way to set them up so I can switch when prices change yet again.
Annual Pro+ subscriber here, mostly using it as a fallback when my Claude Max plan hits limits. The request-based pricing was genuinely the appeal, I could spill over from Max into Copilot without thinking about token costs. Going from 3x to 27x on Opus on annual plans is rough. The newer reasoning models with variable thinking budgets probably made per-request pricing untenable in the long run, but the way this lands on existing annual subscribers feels harsh. Going to look into the prorated refund.
I pay for github copilot solely for completions and use codex for actual agent work. I really wish I could pay like $5/month for just completions or there was a good local alternative for them.
What's the residual value of Copilot after the changes? For Enterprise plans, even copilot code reviews will be charged at token price + github action minutes for the execution of the review. One can roll one's own reviewer and have it spend the API tokens as well. At least one will be able to select the subset of files for the change set, if needed.
The background agents will also depreciate in value because of their harness that's a black box that's not optimized for token usage at all. Rolling one's own will be a better choice here.
Whose idea was this “premium request” model anyway? If you’re going to invent a new metric used to bill, why not align it with what, even at the time, was a clear underlying cost structure that GitHub actively chose to ignore for a more confusing system.
It made more sense in the ye old days where a request was basically just a chat message in a sidebar and it could also edit code. Then saying someone can use 300 chat messages a month kinda makes sense.
Turns out when a request can spawn tens of subagents and use millions of tokens over many turns of toolcalls then suddenly github copilot has a massive financial problem on their hands.
This approach started with the “Ask a question about your code” feature, which is more comparable to single chat message with relatively predictable token usage. Now it’s an agent who might work for 30 minutes, read the whole codebase, and write 1000 lines
I'm not usually a Conspiracy Guy, and the answer is probably `incompetence * tech_debt`. But I think that having sufficient layers of abstraction to any billing model is a useful way to hide the real cost of things. It's why it's done everywhere.
> Code completions and Next Edit suggestions remain included in all plans and do not consume AI Credits.
This is the VSCode autocomplete stuff right? Really enjoy this.
> Copilot code review will also consume GitHub Actions minutes, in addition to GitHub AI Credits. These minutes are billed at the same per-minute rates as other GitHub Actions workflows.
Looks like Microsoft has run out of compute and can't scale it fast enough to serve copilot users and Azure AI Foundry needs, given that the customer base is growing there as well.
I'm happy I invested in local solutions and cutting context to the bone for API providers. Claims about AI being able to fully replace programmers never took into account the long-run equilibrium price of inference.
That's my point. They made those decisions without any consideration for the long run, which would have required them to project the cost of AI services years into the future. Obviously management didn't do that and had no way to do that. It made current earnings look good, though, which was enough for them when they made the decision.
I only use copilot for the occasional auto-complete suggestion. I'm betting I could run a lightweight local LLM with llama.cpp to get similar functionality. Maybe this would be a decent replacement https://github.com/TabbyML/tabby
I was always curious how they can sustain a request based pricing model when requests can range from tiny to huge with all the modalities GH Copilot offers. Was a steal for agentic coding that turned out to be too good to be true, in the end. Still: Thank you for the ride.
I guess this is the "Google App engine" of Vibe coding when google raised pricing of App Engine significantly after being in preview. Doing weird price changes like this when coming out for new services make much more sense than doing this change which will just anger users.
According to `bunx ccusage` I'm easily doing $250-400/day in "real" API costs on my $200/month plan. There's no way everybody else isn't going to do the same thing and completely change the industry again. Both beginner and advanced developers are already hooked on all this stuff and they all know it.
Built credit pricing into my SaaS for AI features and the hardest part wasn't the math, it was that customers can't easily predict their own usage. They underuse and feel cheated, or overuse and churn. Subscriptions hide that volatility from the customer. Usage based pricing makes it their problem, which is honest but harder to sell.
We all knew this from the very beginning but couldn't compete with OpenAI or Anthropic on their subscription-based pricing strategy. It was nuts except for those few corporations burning investor money to keep competition out as long as possible. Now they don't have to hide anymore that subscription pricing won't do it for ai. The pyramid scheme is falling.
I'd be fine with it if they can make Sonnet 4.5 unlimited. I haven't personally seen any major differences between 4.6 or any of the other newer models. Claude 4.5 seems to be the right balance and works great for me.
The problem is that people expect to get the output of 100 people with a $20 subscription by spawning multiple agents. This is unrealistic. I'm using 2 codex plus account and able to manage a repo with 265-300k lines of code.
The point is - if its the same or more expensive per month than a real human employee - why pay for AI ?
Human retain knowledge, product knowledge, can pick up more work often for the same money. And having many of them means your business wont go down if provider suddenly bumps API pricing.
How much is the average pay for a junior developer in US? Its definitely much costlier considering PF and other benefits. Just the math. if you use it efficiently it's much cheaper than hiring a permanent staff. You can maintain a lean team and do all the mundane boilerplate coding with AI.
As a Github Copilot user, who mostly just uses chat in the VS Code editor but still burns through my Pro limit every month -- what's the best alternative price to performance? Claude Code?
"Paid for annual? Tough luck, from now on your usage limits are reduced by 89%. You can do 11% of what you paid for. Good luck if you paid annual a month ago!"
And then they have the gall to say
> "The bottom line: Plan prices aren’t changing"
If anyone lives in a place like Germany or Australia and has an annual sub, please take them to court, you're guaranteed to win because you have reasonable consumer protections and their ToS doesn't stand a chance. 9x reduction is unreasonable and the consumer cannot be expected to see this coming.
In case some diehard enshittifier believes that consumers should know better and businesses should be allowed to get away with it, where is the line? 99% reduction? Is that still okay?
If this situation is to be acceptable then it should be regulated as a financial product like stocks, which come with knowledge tests of "do you know you can lose all of your money?". And come with regulatory compliance and all that.
In a way, this is good news. I am expecting to see a lot less of AI-boosters claiming that they are "productive in ways you don't understand", now that they'll have to dish out upwards of a few thousand bucks per month for whatever PoCs they are building. And the excited hired middle managers and CEOs spending other people's money will have to be much more careful about their budget starting June.
It may already be cheaper if you calculate the hidden cost accumulated by LLM generated tech debt. At least some of it could be avoided by manually doing it better to begin with.
I searched for "Show HN" for the first time ever and I got the impression that most of the costs associated with the projects could have been avoided by never starting them to begin with.
"VibeBrowser", "Financial Database API for Vibe Coders" followed by dozens of self feeding AI projects with barely anyone doing anything that isn't related to AI.
so what's everybody using to get autocompletions in vscode? i've been using copilot just because $10 is cheap, but i use opencode for everything other than completions.
i tried the continue vscode extension, and it seemed kind of janky. are there better options?
I tried antigravity, it worked great for a couple weeks and then Google changed their rate limits and ruined it - I’d get about 3hrs of work before all the ai features just stopped working.
just cancelled mine. Only reason to have it was I could get a lot more done with a single prompt. No reason why I shouldn't go to the model provider directly instead.
This subsidized inference is just a marketing ploy to increase prices and profit.
If common people can have a DIY setup with an open source model cheaper than those behemoths with a scale advantage, it's clear that we have been played.
Time to either self host a Chinese open source model or to just pay the cheap Chinese providers.
Yeah, local is clearly the future. Even beyond the cheap Chinese models you can install the apfel[1] stuff if you're on a mac and want a quick available onboard cli option. And I'm sure people will adapt the Flash-MoE[2] integration to be even better soon as well.
DeepSeek (and other open weight models) can be served by anyone with the means to do it. All of the big open weight models are available on OpenRouter via commodity providers.
cursor, windsurf, and CC are all already on usage-based models so I guess what really matters is whether Copilot's GitHub integration depth justifies the price per token vs the alternatives
I'm not sure I understand this. All I know is now, I pay $39/month (actually less because I paid a year up front), use the agent, mostly on auto--and only choosing a model if it got stuck or in a loop--every day, and haven't hit any limits yet. It seemed to good to be true, after hearing others talk of $300/month bills. I guess it was.
Glad this was announced because I didn't even realize that our (small) team has been paying $20/month for Github Copilot when none of us are using it, so used the opportunity to cancel CoPilot altogether. I think it was free when I first activated it (this was also before Claude, Codex, Gemini, and while not that great, it was why not), and didn't realize it had switched to paid and bundled with our Github bill, and now per usage.
Re 1: Current models don't solve coding. They are useful tool for it though.
Re 2: Open weight models seem to be less than a year behind proprietary ones, so sure, if you're willing to spend tens or hundreds of thousands of dollars on a super computer that you probably don't fully utilize instead of renting time on someone else's super computer for a lot less.
People need to wake up and stop being surprised by these billing increases. I see it on every update of every model. This was all subsidized by VC and company money. Now they need a return and the prices will keep going up. Be glad that you took advantage of that up until now, but can we stop the pearl clutching when we all know the amount of money being dumped into AI and the lackluster returns?
It's less surprise, but more confusing given the game theory as their competitors are not doing the same thing and the multiplier changes alone will likely churn current users.
tldr: people were running multi-hour agentic coding sessions for the same flat fee as a one-liner autocomplete, github was eating the bill, and that party's over on june 1st
Something is hilariously off here: Why should I pay $10 and be forced to use it by the end of the month, while I can pay $10 and have it last as long as I want?
Their "API pricing" is exactly the same as that of providers: https://docs.github.com/en/copilot/reference/copilot-billing...
I'm thinking the same. Downgrade to Pro and use OpenRouter (same price) for overage.
Seems a massive loss for Microsoft. Presumably there's a further rugpull to come.
> Presumably there's a further rugpull to come.
How would that be? They are already charging as much as the underlying providers. They can hardly expect to have any customers if they are charging more.
Enterprise sales will be the answer. Microsoft will have some story that convinces an exec eight levels up the org chart from the normal users that this is an essential product they need to overpay for. Given their existing relationshipsand immense sales team they'll probably have success.
That story is data governance. Corporate already have a data-agreement with MS, storing all their data there. Github copilot is covered by that, while a individual agreement with e.g. anthropic needs lawers involved.
It’s precisely this, and, to be fair, it’s a rational approach given a Data Security Exhibit starts at 6 weeks and can hit 6 months to complete. That being said, I work with regulated data, so YMMV.
Plus if you're in government you have procurement to deal with. You already have an Enterprise deal with Microsoft so you don't have to go through any of that rigmarole.
Can confirm, this is the situation.
Microsoft is simply the default answer for most large corporations. Getting access to some Microsoft subscription is very easy, because of the existing framework agreements, Microsoft providing any and all compliance slopuments needed and already being pre-cleared for corporate data etc. Meanwhile trying to use another provider (e.g. Anthropic) would be a one year endeavor, minimum.
We also pay $300/month for Azure Desktop VMs.
We are paying for tens of thousands of those machines, although everyone knows they are stupidly expensive and incredibly slow.
They list the price 900% higher and give a 90% discount to enterprises who also use teams, outlook, office or even windows if they're desperate. Then that becomes a deal so good that enterprises can't afford not to take it!
I'm already on Pro. Why should I keep it?
e.g. if on an annual plan? 0x will be gone, but there are okay 1x and 0.3x models left. I am pretty much curious how the early may test invoicing will look like. current setup of tools etc. is way too chatty eats up 1+M token per PRU easily. not sure how much is cached.
Only reason to keep it is if you like their UX and auto-complete. Everything else is on pay per use and if you don't use all of it (good luck with the 5 hour and week caps) you have just paid more for the auto-complete
The deal is really pretty much garbage now and I believe that is the intent.
I use auto-complete mostly, so I'm somewhat relieved. When I do need to use the agent, I don't think I will use all of the tokens.
$10 a month for auto-complete on a good UX is good value IMHO.
private repos
Can you elaborate please? I still clone my private repos and work on them using OpenRouter Credits and OpenCode (or Copilot itself: It supports OpenRouter BYOK). No?
Copilot pro, not github pro
Also you can have private repos on GitHub Free.
OpenRouter charges a 5% (?) fee for buying credits.
A very reasonable and fair markup that is clear and well articulated and not changing (at least so far) on a whim.
Yeah, but you get the benefit of using any model of your choice.
is $10 Pro monthly subscription a pre-requisite before i can purchase $10 in API credits?
PS: i would have loved if I can directly buy $10 in credits and be free to spend it as quickly or as leisurly as I want -- without any monthly expiry or fixed recurring payments
I have to wonder if it's because of how many Enterprise customers they have who have standardized on Github Copilot and gotten it through the gauntlet of legal approvals etc.
This rings true to me as someone who's worked at a few large corps like this. A price hike does not change things when there is a mandate to use MS products over other vendors.
That is exactly where I am.
We're putting other providers through the gauntlet. An M4 Studio or two running the latest Qwen3 or whatever counts for state of the art in open models is also looking a little more viable all the time.
They can be super complementary! Open weight models can be your everyday standard goto, and frontier models for the harder and bigger tasks.
Having some open weight deployment or vendor is also a good thing, because you may have domain specific tasks where you can get better results on domain specific problems with a quick finetune.
Unsloth makes it particularly easy. Open weight LLMs are incredibly powerful building blocks.
Bingo. Ghcp is the only allowed llm solution at our big well known semiconductor corp. It'll take years to get approval for anything else. We're stuck with it and will pay whatever price we have to.
I'm wondering if they're basically saying they're going to give $10/month free API credits to students and open source maintainers and so on... while otherwise getting out of the consumer portion of this space.
they're downsizing free github copilot pro for open source maintainers. At the very least, it looks like small open source projects got their free copilot pro cut off
I was kinda of disappointed when that happened earlier this month, but not as much now after seeing this change. My primary use had been trying out some of the newer Anthropic and OpenAI models, which probably would have burned through $10 worth of credits rather quickly given their new pricing.
I am a bit confused by the separation between VSCode and Copilot. If I cancel my Pro+ subscription, can I still use Copilot with my own OpenRouter key?
for my experience currently, I greatly prefer the VSCode Copilot extension experience over the Claude Extension
I think VSCode only supports copilot for "autocomplete" too
on top of that, you need GitHub Copilot for the PR reviewer functionality in GitHub
Huh, I find my copilot plugin to be so incredibly glitchy. My agents are always reporting that their shells are mangled that commands are truncated and all kinds of nonsense. Sometimes they spin up dev servers fine other times it just hangs waiting for a terminal response. So far I have found relying on the CLI from the model providers to be significantly more reliable.
I do like the integrations with the IDE however, they are convenient for rapidly reviewing changes. I just need their terminals to actually work!
I had this problem and it turns out it was my oh-my-posh command prompt customization. VS Code injects certain control characters into the output stream for agents to observe events and the theming runs after those mechanics are hooked up so it can interfere. Updating to the latest oh-my-posh fixed it for me.
Here's the oh-my-posh GH issue[0] in case your problem is similar but not solvable with a simple package update.
[0]: https://github.com/JanDeDobbeleer/oh-my-posh/issues/7029
Ive also had huge problems with copilot and terminals, mostly due to corporate antivirus which makes it impossible to install clink for example. So the Shell integration never really worked. It has been fixed for a couple of months though now, and its working, maybe not great but at least very good.
The only sad thing is trying to use tools in a VS developer prompt (and how could this not have been fixed ages ago, its literall YOUR OWN flagship product). It knows how to launch the .cmd for it, but thats incredibly slow for single commands. Would be nice if I could tell it to just use an open terminal.
I have the same problem with zsh + powerlevel10k
but surely the issue is on VS Code side, to do things in a way that work with people's shells as they are
other agent harnesses don't have the same problems with my shell
You can use Copilot extension with OpenRouter (among others).
And yes, I need to find a solution for autocomplete. It used to be available in free tier of Copilot. Not sure anymore.
Enterprise gets pooled credits and will like having everything go through one place so I think it still works.
You can pool credits through open router (afaik, I'm only using a single user account), but if you top-up $10 per user, per month, any unused credits will rollover.
Tbh I think it still works, but only because the new allowance will likely get used very quickly within a billing cycle - I'm expecting this change to increase our orgs bill significantly based on how many API credits with open router I consume in a weekend using a single agent in a pairing style.
The pooling will only be useful if you have a bunch of infrequent/low usage users that you still want to have licenses.
Which is almost guaranteed to be the case for a large org, considering everyone will want auto complete and PR reviews, but on average most will not be making a ton of agent use
kilocode allows you to keep your credits at the end of the month, and if you run out, give you an extra 50% I think.
(No affiliation, they're my next stop when my trial of copilot runs out).
Is there a way to use the autocomplete feature with an api?
No but autocomplete is not part of this billing change
Is that autocomplete better than IntelliJ own plus their local only LLM completion?
I uninstalled copilot plugin because it was eating memory and its completions where about 60% good and the rest was bad.
After switching back to IntelliJ I see just positives.
The era of subsidised inference is truly ending. The new model multipliers (https://docs.github.com/en/copilot/reference/copilot-billing...) seem like a huge leap, though. From 1x to 6x for new-ish GPT and Sonnet models. 27x for Opus...
Seems like folks would be better off with OpenRouter instead.
Lots of us have noticed that usage limits for Claude have been nerfed in recent weeks/months.
If anything, these new multipliers are more transparent than anything OpenAI or Anthropic have communicated regarding actual costs and give us a more realistic understanding of what it's costing these providers.
The fact that we were able to get such a substantial amount of usage for $20/$100/$200 a month was never meant to last and to think otherwise was perhaps a bit naive.
This feels like a strategy from the ZIRP era of tech growth where companies burned investor capital and gave away their products and services for free (or subsidized them heavily) in order to prioritize user acquisition initially. Then once they'd gained enough traction and stickiness they'd then implement a monetization strategy to capitalize on said user base.
However, inference costs for entirely good enough models are likely to keep declining in the future. We're probably hitting diminishing returns on model size and training. The new generations aren't quantum leaps anymore, and newer generations of open source models like DeepSeek are likely to start getting good enough.
There's going to be a limit to how much they can raise prices, because someone can always build out a datacenter and fill it up with open source DeepSeek inference and undercut your prices by 10x while still making a very good ROI--and that's a business model right there. Right now I'm sure there's a lot of people who will protest that they couldn't do their jobs with lesser models, but as time goes on that will get less and less. Already right now the consumers who are using AI for writing presentations, cooking recipe generation and ELI5 answers for common things, aren't going to be missing much from a lesser model. That'll actually only start to get cheaper over time.
Also for business needs, as AI inference costs escalate there comes a point where businesses rediscover human intelligence again, and start hiring/training people to do more work to use lesser models--if that is more productive in the end than shelling out large amounts of cash for inference on the latest models. [Although given how much companies waste on AWS, there's a lot of tolerance for overspending in corporations...]
> because someone can always build out a datacenter and fill it up with open source DeepSeek inference and undercut your prices by 10x while still making a very good ROI-
Not sure how it all works out. Currently trillion dollar companies can't make a native app for platforms. Everything is just JS/Electron because economics does not work for them.
And here companies can make GW data center running very expensive GPUs for 1/10th of current prices. Sound little fanciful to me.
The price you pay for anthropic must include the price of training new and better models which is incredibly costly. If you use the models someone else already spend money to develop you don’t need to pay this price.
I guess the new models will still be quantum leaps, but literally: "The smallest possible change in a system"
They've been like that for a while actually, I think at least since the big hype around ChatGPT 4.5 (or was it 5?) and that underwhelming, lukewarm, oversanitised presentation by Altman and his team.
Yups... Mythos is the smallest possible leap. Not a standard model generation advance, not even a version point advance. Just the smallest possible quanta of a change. We are absolutely hitting a plateau any day now. Any day. Any time. Any second now. Yup. Right now! Surely!
I mean let's be realistic - all that we know about the "mythical" Mythos is the carefully curated and release stuff by the Anthropic's PR team. Is it really a huge leap they are making it to be? I doubt it. In fact I bet if it was indeed that powerful and dangerous, as they imply, they'd find a way to release it immediately, devastate OpenAI and DeepSeek and secure a leading position in the market. Why is it not happening? I suspect because Dario is again at it, peddling his bullshit.
Yeah. AI progress is insanely fast if you compare it to anything else. Where else is a one year old technology already hopelessly outdated? 10 years ago is basically stone age.
I am continually tripped out by the fact when I was 16, I didn't have a 'smartphone' beyond a Windows Mobile 6 phone that had no internet on it.
Now, I have this high-resolution shiny object that can near instantaneously get any information I want along with _streaming HD video to it_ *anywhere*.
15 years even feels like a stone age. I can't fathom what it has to feel like people in their 60s and 70s.
I'm not quite 60, but it's always interesting to me that I feel quite the opposite of this. When I was 16, I didn't have a computer, didn't have a phone, had never used the Internet, but when I think of how life has changed, it's frankly not much. I woke up this morning, scooped my cats' litter boxes, took out some trash, made myself breakfast, ate that, read some news while eating, then lifted weights in my garage, had some work meetings, wrote up some instructions per a customer request from Friday, and am about to go drive to the lake to go do a 9 mile longboard loop.
That's very close to a normal day in 1996. The biggest difference is I read the news on my phone instead of a physical newspaper. The news was not any more interesting or informative because of that. I guess I can also still do the loop reasonably well, but I'm a lot slower than I was in 1996 when I was a cross-country state champion.
My parents are closing in on 70 and I guess I can't speak for them, but I'm at least aware of the daily routines of their lives, too. Walk the dog, do housework, DIY building projects, visit kids and grankids. Seems much the same, too, with the biggest difference being they're now teaching my sister's sons to play baseball rather than me, but shit, one of her sons even looks like exactly the same way I looked when I was 7! The more things change, the more they stay the same.
General agree... I still do the things (mid-50's) I used to do when I was a teenager with no computer, no phone.
But - now they are easier - I can read books on an e-ink screen and pretty much instantly find what I want to read next. I get my news on a phone. I used to watch TV/movies broadcast or on tape rentals. Now, I have just about everything I could ever want available - without ADs... those were such a time-waster.
What has changed is that I have access to MORE information than my local (or school) libraries could ever provide - in a variety of more accessible formats. Whatever tools I need to get "work done", I can find a myriad of free and open-source options.
But - the overall days and household family routines are the same - now, instead of reading a paper book while waiting to pickup my kids (or other family members) "back-in-the-day", I can read my device, or connect with my DIY communities online on my phone - or learn something new. I don't have to schedule life around major broadcast events, I can easily do many tasks while I am "out-and-about".
Friction has been reduced.
If your parents are closing in on 70, I would have expected you to be closer to not quite 50 than not quite 60.
I am just over 50 myself and I agree with your points. Technology has changed but life is largely very similar to wear it was in the 90s. At least day to day. Attitudes are way worse now.
Thank you for this insight!
I always wonder the views of older people. My parents are very technology forward and have been my entire life so it is difficult to gauge how different life is compared to when they were growing up.
It's easy to hear "Oh well I only had 640kb of memory and typed programs out of a magazine I got in the mail!" and see as distinct from having 'unlimited' resources and the internet.
Your insight is good ("The biggest difference is I read the news on my phone instead of a physical newspaper") that life sort of stays the same but the modality changes. People still go to the store like they did in the mid-1800s but now it is by car.
I wonder what our "industrial revolution" will be where the previous generation lived (ie: out in the country on a farm) totally different lives to the current (ie: in the city in a factory). Maybe when space travel and multi-planetary living is normalized?
> It's easy to hear "Oh well I only had 640kb of memory and typed programs out of a magazine I got in the mail!"
Since I was there (young, but there), I want to point out that this crosses three eras which all felt quite different:
Exact years vary but these eras were only about 5 years each. Nobody had a floppy in 1978 but almost computer user did by 1983; nobody had a hard drive in 1983 but almost everyone did by 1988.To some degree this already happened with the move from the industrial city to suburbanization and then re-urbanization. In particular one of the most notable recent developments is that urban waterways are now pretty desirable places to be with parks and recreation; in most industrializing cities the waterfront was actively avoided because the industrial use made it polluted, smelly etc.
Depends on where you live. My dad is almost 80, grew up in a very rural area, and when he was 16 they'd just gotten indoor plumbing. Up until he was 14, his school was a one-room school house with no heating other than a wood stove. If you were the first kid to arrive for the day, it was your job to get the fire going in winter months.
Housework meant no laundry machine, no dishwasher, and possibly no vacuum cleaner. That means hand washing everything, and beating rugs with sticks and brushes to get the dust off of them.
The news on the phone is worse, in fact.
I think so too.
And at some point even frontier model costs will hopefully come down (if there is still a meaningful difference between closed and open source models at that point) as all of the compute that's being built out right now comes online.
I hope it's true, but right now hardware prices are insane
It does feel like the music is about to stop.
It has been years now, of cash injections, investors can't keep feeding the beast forever.
This is the best AI programming will be. From here on the enshitification starts and the prices go up.
As predicted by many. The math is, as usual, mathing.
It has been years now of reading this same comment... Surely people can't keep typing it forever.
But the prices haven't been going up by multiples of 6 for the past few years. Things are actually changing now. I don't think it's over, but in the short term, it's going to be considerably more expensive.
They will smooth up the spike. Or be subtle and transform the existing quota so that they run out more quickly. Calling it caching, compression, optimisation, of course for the sacred benefit of the users.
That would be, even is, the smart thing to do.
And it didn't really get flawless, did it? All the same objections stand, but the cost is inevitably blowing up for the same kinda jank product.
The difference is we're now in a world where Disney has pulled out of OpenAI without comming, and Sora was dropped off a ditch.
In other words.
The bubble has burst. You're just in denial.
I’m not willing too, but I can set up a cron job to Claude -p the task.
Dunno, if in this day and age you are making inference more expensive, more scarce, you are honestly moving in the wrong direction and DeepSeek and others will gladly take your lunch.
The hardware to run deepseek is still incredibly expensive.
> The hardware to run deepseek is still incredibly expensive.
Deepseek API pricing is very low compared to Anthropic/OpenAI API pricing.
For many, the 300% difference in pricing may be difficult to justify, if the quality difference is very small. And there will be many tasks where the most expensive/the best model, is not needed. Currently many people end up using Opus 4.7/GPT 5.5 for many tasks without thinking about it.
Is deepseek still on subsidized pricing though.
Near zero probability of that. The model is more efficient and the company who trained it did not blunder trillions of dollars to do so. China has better electricity infrastructure than the US too, so the likelihood they can scale out before the US ever could is high. Long term deepseek, Alibaba, etc hold the most cards for sustainable AI even despite the attempted Nvidia embargo
I am not shilling China, this is just what is happening right now.
Lol what? You seriously think that the #1 Chinese AI company is not being subsidized by the Chinese government?
Which one is #1? Alibaba, deepseek, or moonshot?
I think the Chinese government works differently than the US government. I think China has been subsidizing their electricity grid for decades and leading the world on sustainable electricity namely solar. While the us has let their infrastructure rot and laughed at government inefficiencies for about half that time. The US has data centers running on gas right now while waging wars blowing up gas infrastructure world wide. It would be comical if it wasn't an environmental disaster. Most of them have no hopes at even getting enough power in well established areas short term.
I realize what I am saying may come off as propaganda because the US holds net negative views on China so here are some links.
https://www.technologyreview.com/2025/07/10/1119941/china-en...
https://www.wired.com/story/data-centers-are-driving-a-us-ga...
I think because openai spent so much money upfront showing how it was possible to do this and laid out a product roadmap China got to get on board much cheaper and easier. I see no reason to not believe any of these companies when they say they didn't squander tons of money to do what they did because I don't know how openai has even spent all the money they have it's actually ridiculous to think about.
https://the-decoder.com/openai-adds-111-billion-to-its-cash-...
Judging by the multiple providers selling it for around the same price (including non-VC funded competitors): no, it isn't subsidized.
Is there somewhere I can look up whether a certain provider is VC-funded?
It's not really about that. China is eating the US's lunch when it comes to ai. Don't get me wrong opus is the strongest model out there today, but that's the us's only advantage right now. Deepseek,qwen,kimi, etc all have fundamental research making the models smaller, more efficient, scalable, etc. in the US the plan is to buy all the hardware, write legislature, embargo other countries, keep models and research closed, so people cannot innovate for the next two to five years.
Unlike the us chinas focus is on research and sustainable building. China also has really good infrastructure for energy, etc. it is also to their advantage to drop 5 billion instead of 2 trillion and beat the us while turning a profit.
Chinas focus in ai is less flashy and because they are the biggest manufacturing super power in the world right now, it directly feeds their economy. They aren't looking for applications or to replace thought workers with slop bots, they have natural needs for this technology. Us manufacturers can't compete so they have to keep companies from selling their goods there see byd. China sees it as commoditizing their complement, the us is risking its entire economy and it's environment and resources, kind of scary.
Have you seen the news about qwen3.6? People are running it on sub 1000 euro hardware. Apparently it's about as good as Claude sonnet.
That is folly because there is very minimal cost to switching providers, let alone models.
Did anyone really expect AI to be cheap?
If/when it gets to the point where it can replace a skilled worker, the service can be sold for close to the same price as that skilled labour. But the AI can run 24/7, reliably, and scale up/down at a moments notice.
There's not going to be much competition to drive prices down, the barriers to entry are already huge. There'll likely to be one clear winner, becoming a near-monopoly, or maybe we'll get a duopoly at best.
> Did anyone really expect AI to be cheap?
Yes, a lot of people (not me). Why? Well because that was the whole value proposition of these companies, relentlessly pushed by their PR and most of the media- rememmber it was something something Pocket PhDs, massive unemployment etc?
"There's not going to be much competition to drive prices down, the barriers to entry are already huge. There'll likely to be one clear winner, becoming a near-monopoly, or maybe we'll get a duopoly at best."
Based on what exactly? So far every time OpenAI, Anthropic or whatever has released a new top performing model, competitors have caught up quickly. Open source models have greatly improved as well.
I expect AI to be just like cloud computing in general - AWS, Azure, GCP being the main providers, with dozens of smaller competitors offering similar services as well.
Right now China is flexing the future in my opinion. Smaller, widely available, frontier models for pennies on the dollar.
I think the future of ai will be breakthroughs that let it run on commodity hardware, and the average person will not be paying for it from the cloud unless they want to be surveilled or are stuck on older hardware.
Right now I am running about what was a frontier model 1-2 years ago on a junk machine. Some people are running what was a frontier model 4 months ago on PCs and laptops that cost 5,000. In a year I think the landscape will be even better.
I do. "Commoditize your complement". Want to sell lots of silicon? Give away good local models to run on that silicon.
Even if SOTA models in the cloud are a few percentage points better, most work can be routed to local models most of the time. That leaves the cloud providers fighting over the most computationally intensive tasks. In the long term, I think models are going to be local-first.
(Unless providers can figure out a network effect that local models can't replicate).
> I think models are going to be local-first.
Why on earth would that happen when everything else is moving into the cloud to tie it to ever-escalating subscription fees and prevent piracy?
Even with gaming, where running high-end 3D games in the cloud seems like madness and inevitably degrades the quality of the experience, they won't stop trying.
> In the long term, I think models are going to be local-first.
Why? There's an inherent efficiency advantage to scale, while the only real advantage for local models (privacy/secrecy) hasn't proven convincing for broader IT either.
It's foolish not to care about privacy especially as a company. You know how it prevents you from emailing yourself your tax documents? Meanwhile thousands of employees are sending literal design docs, software, product goals, etc to several ai third partys. Not only is that insane, the companies they are sending it too intend too and openly admit to scanning the data, make software products themselves, and intend to create models that can produce their products automatically.
The reason local models hasn't caught on is several fold. It's marketing to say your company follows the latest trend, and there's an inherent pressure to keep AI companies afloat so the economy doesn't entirely collapse. The other is, it wasn't until the last month that these models have caught up to frontier models. They just did, and they are more efficient and don't require a team of 500 to deploy.
Local first models aren't just more private than the API vendors, they also have the advantages of fixed cost, lower latency, and better stability - local models don't get nerfed/"updated" in the background like chatgpt does.
Maybe in a world where these AI companies behaved with some semblance of ethics and user-friendliness they would be on even ground, but for anyone paying attention local models are obviously the future.
> the only real advantage for local models (privacy/secrecy) hasn't proven convincing for broader IT either
Because of nonexistent regulation. Just wait for it…
The legal situation in for example the EU is crystal clear, only that it will take some time to go though all court instances.
To not depend on an external company that can decide the price.
That's a silly reason. For non-agent use cases what kind of utilization are you going to average on your own GPU, 5-10%? And that's without batching.
Even with overhead and scaling for peak use and a large profit margin, any company with an ounce of competition will be vastly cheaper than self-hosting. And for models you can run yourself, there will be plenty of competition.
I think you are calculating with current prices. Try to extrapolate the price in one year, seeing the current trends instead.
Extrapolating current trends, I expect API prices to drop significantly for a given measure of 'intelligence'.
> Did anyone really expect AI to be cheap?
Considering most of the cost of producing a model is the upfront cost rather than the running one, I kinda still do.
The point was never to produce 4 frontier models per company a year.
"This change aligns Copilot pricing with actual usage and is an important step toward a sustainable, reliable Copilot business and experience for all users."
I see statements like this as strong indicators that the sales people are wrapping up their work and the accountants are taking over. The land rush is switching to an operational efficiency play.
The sooner the better. Let's take a look at the long term, enshittified, viable product before we get too dependent on the trial version.
And enshitification starts.
Yeah, totally. The recent pricing changes have just made my Copilot subscription go from great deal to awful value over night.
I've been wanting to get off MS more generally and this is good motivation. Will be playing round with OR this week.
Just be aware OpenRouter charges a 5.5% fee, I didn’t know until recently. I like the product, and I think the fee is fair, but if you want the absolute best pricing then go direct.
But with open router you can always just use the latest model. If you're committed to eg Claude opus then you're better off going directly to anthropic for sure, but if not, varying other models may be fine too, depending on use case and be massively cheaper. Eg new deep seek model with same mio context window or Kimi k2.6 with 270k context window for subagents which implement
>but if not, varying other models may be fine too, depending on use case and be massively cheaper
Do inference providers have standardized endpoints, or at least endpoints compatible with claude code? Otherwise to pay 5.5% on all your tokens just so it's slightly easier to swap providers (ie. changing a few urls?)
> Do inference providers have standardized endpoints, or at least endpoints compatible with claude code?
Yep, you can plug deepseek/kimi/minimax into claude code just fine. Or run everything through another harness like opencode instead.
Or you could use gcp Vertex or aws Bedrock and still have access to a bunch of FMs without a markup.
Wow thats a lot for routing traffic.
And handling API tokens, and billing, and reliability, and middleware. I am not affiliated with them but it’s not “just” routing.
Apple still charges 30%. 5.5 seems pretty reasonable. /shrug I dunno.
> handling API tokens
Don't you still need to handle tokens with them? Also that's trivial.
> billing
Yes but you'd be paying for billing anyway.
> reliability
They increase reliability?
> middleware
Which you wouldn't need if you paid directly.
I'm not saying they shouldn't get 5.5%, but that list is mostly non-convincing.
> Apple still charges 30%.
3 of the 30 is for billing, with the rest mostly being gatekeeping with a fake justification on top.
There's nothing trivial about getting a Google API key. Openrouter removes that stress from my life. And I can route requests to providers above a certain TPS threshold. And much more.
Payment processing likely eats up at least 2-3% of that
IIRC OpenRouter charges you for the payment processing fee also.
Still worth it IMO to be able to switch from Provider A to Provider B if Provider A is having a bad day.
I will not be renewing/switching over, either.
I had copilot mainly so I could write issues and throw agents at it, while I went off and did other things. Has been great for contained spot work.
At this point, I'll go ahead and leave it expire, and then consolidate between Codex and JetBrains AI. Especially since Xcode supports Codex with a first-party integration.
Even Sonnet 4.6 is 9x multiplier (previously 1x)!
The only model I even used on Copilot was Sonnet and now its got a ridiculous multiplier.
At this point they might as well just charge per Million tokens like every other provider instead of having a subscription.
They do for any new plan. Those multipliers are only for people that paid annually. After their subscription ends they'll go into token based pricing like the rest of people.
I understand it like : the 10 usd is for handling the business record, maybe also the harness, I get a few coins to kick tires, but to use it for anything real it’s pay as you go by the tokens list price.
> At this point they might as well just charge per Million tokens like every other provider instead of having a subscription.
Pretty sure that's what they will eventually do
... that is exactly what they will do. Just click the link in this thread, or read the headline.
Why the multipliers then at all?
The multipliers are there only for current annual plan customers. After 2026 its all tokens.
I thought I was smart for buying the annual plan after I graduated and lost my student plan and then GitHub taking away my Copilot Pro I got for free for being a author of a popular OSS project. Turns out I'm being punished for making that year commitment to them. I like to think I'm only a moderate user of GHCP so this is just terrible for me. I'm honestly thinking about cancelling and switching to alternatives while also looking at investing in a local LLM setup.
So they're changing the product that people already paid an annual subscription for to the worse. That's asking for legal complaints.
27x for Opus is genuinely shocking. at that point you're not paying for convenience anymore, you're just paying a GitHub tax. OpenRouter or direct API makes way more sense unless you're really glued to the IDE integration.
I keep seeing people mention OpenRouter.
Does it effectively bypass regional restrictions for you, so you can use something like the Claude API from unsupported regions such as Hong Kong, or does it still enforce the official providers' geo-restrictions?
OpenRouter is great for budget control, but as they are indirect APIs, your experience with cached tokens may vary, eventually costing much more than in direct depending on the providers.
You can pay with crypto though, which seems to be convenient for people under sanctions or with limited access, or if you are in low-tax jurisdiction (e.g. HK)
Caching is advertised per model+provider.
That said I think few people using openrouter are actually being selective about providers.
It took half a day to get my opencode setup, was not friendly. A lot of manually cross referencing model and providers. I was actually mainly optimizing for relatively fast providers. It all is super fragile and I'm sure half out of date; I have no idea if these picks are still fast, no promises they are still the same price (pretty terrifying honestly).
I'm mostly on coding plans so it doesn't super affect me. But man is it a bother to maintain.
Wow, having a corp. account I do wonder WHEN we are getting some kind of resctriction of usage, or require us to justify our usage.
That GPT4-mini change is going to be brutal! Its much better than 5-mini, which was itself much better than earlier free models.
The point of this loss leading is to properly hoover up the money in the pockets of enterprise customers, get them locked into the idea that they need the latest and greatest cloud-based model, while simultaneously starving everyone of the memory they'd need in order to run competent models locally.
In not-too-distant future we're going to be running better models on our phones than we can buy access to today in the cloud. Skate where the puck is going: soak the customers until that day comes.
It's interesting that the cost multiplier for Claude Sonnet 4/4.5/4.6 varies so much (1/6/9), while the API cost is exactly the same for all three models.
Also, the multiplier of 27 for Claude Opus 4.6/4. is way higher than the increase in API price would suggest.
I wonder why that is.
On GitHub copilot you pay per prompt. More powerful models can do a lot more work (consuming a lot more tokens) per prompt. Also, they tend to use more thinking tokens.
> More powerful models can do a lot more work (consuming a lot more tokens) per prompt.
That is not my experience. Each model since at least GPT-4 can fill up an entire context window. In fact, more powerful models can solve tasks faster, so their ratio of multiplier to API price should decrease, not increase.
For example, Claude Sonnet 4.6 has a multiplier of 9 and an API price of $15, which is 0.6 multiplier per dollar.
Claude Opus 4.7 has an API price of $25, so it should have a multiplier of 25 * 0.6 = 15 when extrapolating from Sonnet, but the multiplier is 27.
> Also, they tend to use more thinking tokens.
That might be it. Is there any data on this somewhere?
Those multiplier are only for grandfathered Pro an Pro+ plans that had annual billing, basically a way to scare people of out of those plans. Ant new ones (and bussiness+enterprise plans) will be on token based billing since June 1.
One theory of the play of SpaceX might do if everyone migrates to query-based billing:
Provide cheap and unlimited access to Grok for programmers (hence the Cursor partnership/purchase for distribution).
-> This would drag massive revenue right before the IPO announcement, like if the company is super growing
-> At a loss, but don't worry, we need these funds to build the biggest datacenter of the universe.
This announcement would create enough momentum to increase valuation, and because of the merge of his companies, would save his X/Twitter investors from a tragedy.
-> Would also be a great service to Cursor investors and so, who are stuck with their VSCode fork
They probably want the training data. Otherwise these 60B don't make sense at all.
But they can't buy curser before their IPO so thats that?
Perhaps they have to much compute because Musk overpromised and Twittergroq doesn't need that much compute after he nerved the porn stuff?
It takes longer to build a datacenter with that much capacity than it does for the market to respond.
Buying real estate in imaginary places is lucrative at first
I think they're going to have to do a lot to overcome the Musk and Grok poison. Even ChatGPT hasn't had as many lapses as Grok has had.
I don’t get the SpaceX reference. I thought they made rockets?
Nobody is paying for Elons xAI so he used SpaceX to buy xAI to fund it.
Under the pretense that SpaceX will be used to launch material into space to build space data centers.
They now also own xAI
Which in turn owns Twitter. SpaceX is now a social media company in addition to a rocket company.
One theory I think Matt Levine posited, is that SpaceX will go public with dual-class stock that gives Elon control even with a minority ownership stake, and will subsequently buy Tesla, which doesn't have dual class stock, making SpaceX the singular "Elon Musk company", with him having operational control despite being public.
That theory aligns with Elon's long-held dream of X as the "Everything Company".
Then he'll rebrand SpaceX to " X" (a space followed by an X).
Can't wait for people to migrate to open tools (opencode/openrouter). This will unlock a lot of innovation.
(I know openrouter is not open, but it allows competition and should be easily replaceable if needed)
Show HN timing matters more than people think. Monday-Thursday, 9-11am Pacific, is when the front page has the most engaged readers. Weekend posts get less competition but also less engagement.
Why would folks be better paying 5.5% fee to OpenRouter ("Open") if most people just use one or two providers? Just use the provider's API.
The routing automatically routes you to other inference providers (for the same model) if/when the original provider goes down.
It's a convenience cost, for sure, but it's not valueless in a fast-moving world. Certainly if you're comfortable with one provider and it's cheaper, do that.
For me the largest value-add is the unified API. Being able to instantly start trialling a new model with zero code changes is well worth 5%. The other part is not having to deal with billing for multiple platforms.
What's annoying is that it's obvious. In the case of GPT 5.5, if Copilot is going to charge 7.5x what GPT 5.4 costs while OpenAI themselves via the API/Codex only charges 2x of what GPT 5.4 costs, that will immediately raise an eyebrow.
To anybody who's been watching the tech sector with a critical eye for pretty much any period from the late 90s and onward, this is just the enshittification process. For most of OpenAI's existence it's been obvious, to me, that investors were burning insane levels of capital to build the market, and now that folks are locked in, you're seeing higher fees, ads, etc. Yet again, the user is the product; the investors want to siphon your data, attention and once you're hooked, money. And for companies like Microsoft and Apple, those hooks can dig deep.
Let's call it for what it is dumping. Dumping things on market below cost of production. This should not have ever been allowed. RnD costs I can accept somehow. But in this case the interference should have always been billed for the real costs that it took to produce and pay off the capex.
I'd call it a straight up "bait and switch".
If you paid attention to the power requirements and amount of hardware being put into data centers, you should have realized that it cost them an order of magnitude more than you were being charged. To rework your analogy: they hooked you, now they're gonna see if they can reel you in.
They can only reel you in if its worth it. I still can code.
And while i do not spend 200$ privat, in my startup we discussed this and our current mental model is, that instead of hiring someone new, we prefer to have more money for tokens.
This is easier for us and has a bigger benefit. The cost of a new / first employee is very high, a 200$ subscription is not. Upgrading that to lets say 400 or 800$ is still alot easier and if i can run multiply and better agents with that money, lets goooo.
I'm looking at education -- teachers and students, not terribly tech savvy, are being mandated to use these tools. And then comes the rug-pull. It was worth it, but now it's outside of their budget. Poorer schools / students can't stay at the cutting edge; richer schools / students can.
You still get far with 20$ if you don't use it daily for lots of coding and thinking though.
And Gemma 4 and other open models can easily be hosted even for schools.
Oh, I thought it was opium.
“Enshitification” is just when unsustainable subsidies end?
Another reason to hate that word.
From a different perspective, you were granted an incredible gift from the companies who let you use their product on their dime. Hopefully you made the most of it when you had the opportunity.
No, it's much more than that. It starts with unsustainable subsidies, as Uber undermined the taxi industry with a ludicrous burn rate. And then, once everybody's hooked to the point that they can't imagine life without the product, you raise costs. And you iterate: raising costs, lowering quality, selling data, increasing addictiveness. Until everybody wants to get rid of it, hates every aspect of it, but is still hooked to the core product. I'm personally not using these tools, not using uber or Meta products. But I'm still using some Google products and it's hard to extricate them from my life now that I'm using them.
> No, it's much more than that.
Okay then this AI stuff isn’t an example of that even under your definition.
unless the 5.4 price is a huge loss leader for them
Everyone seems to believe OpenRouter isn't subsidizing but, until they publish audited financials, I personally doubt it.
OpenRouter doesn't even have hardware. What are they possibly subsidizing? The platform costs?
OpenRouter is guaranteed to be about the highest margin operator in the business right now. Everyone wishes they'd be them, skimming 5% off as the middleman without any OpEx.
> OpenRouter is guaranteed to be about the highest margin operator in the business right now. Everyone wishes they'd be them, skimming 5% off as the middleman without any OpEx.
The 5% fee probably has to factor in Stripe's fees, which would be around 3% to 4% depending on whether it's an international card.
Streaming, caching, and tool calling can get pretty expensive with scale, even when you don't touch inference. Maybe they're doing something clever and are quite profitable.. or maybe they've already taken $40mm from VCs and are currently trying to raise $120mm at a 1.3B evaluation.
They also show headline prices for the cheapest provider of whatever model, but then need to hit different backends some of which may be more expensive. For now they absorb those costs, but the VCs always come knocking.
Just my opinion though. Totally agreed that they have one of the best positions amongst all AI providers from a financial standpoint.
> They also show headline prices for the cheapest provider of whatever model, but then need to hit different backends some of which may be more expensive. For now they absorb those costs, [..]
They do?? I was under the impression I was just playing the price for whatever provider they deemed 'best' for each completion.
That is what I had heard.
Checking now: The way they describe it in their FAQ is that if the price changes, then they will bill you the new price. But I read that as regarding if the primary model provider changes their headline token cost; not in the case of pricing differences for models that have many different backends that host them.
Regardless, I would be more concerned about the streaming costs if the service continues to blow up and they scale aggressively through VC investments. If their 5.5% skim accounted for what they needed, you'd think they could effectively grow organically..
FYI, these are the multipliers for annual plan. I would hazard a guess most people are not on an annual plan
I am and I see it as stopping the music at a party when you want everyone to go home without telling them to go home. There is also the offer to quit with prorated refund for the remaining time. I think I am going to take it.
We can't even get slop delivery worked out. So use SlopAggregator instead.
I don't know if it's just me but copilot kind of sucks. I've been running local models with like 9b parameters and they are about as good if not better. Obviously there's no integrations or whatever and I get most people are probably paying for that than anything else but eh. Big no thanks from me.
"eras" tend to not be so short lol
That's so unfair to us hard working developers. A month ago i could buy for .4$ a turn with Sonnet. Now i have to pay at least .9$ for this turn. Weeks ago i could buy for .12$ an Opus turn after they already raised prices and now they want .27$ from me for the same product! They are stealing from us!
They aren't stealing from us, for several reasons. First of all, it's a voluntary transaction. If you don't like the prices, use something else. Or don't use AI at all.
Second, you have no idea what their costs are. It is most likely that they are simply passing on their costs to you. If that was not the setup, users would just go to another service provider who was providing tokens at a cheaper rate. It's not like there is a dearth of competitors in this business.
The already stole when they trained their models on the data.
Now they just increase the price to buy it back
"Your plan pricing is unchanged: Copilot Pro remains $10/month and Pro+ remains $39/month, and each includes $10 and $39 in monthly AI Credits, respectively."
If there's no discount on credits (in terms of tokens per dollar) over other providers, I'm going to switch to a PAYG provider. If there's a month where there's little to no coding I can pocket the 10$. What incentive do they give to stay with this plan?
Yep.
Or if you're a business with multiple seats, these plans may be more inefficient than raw API usage billing. Since if anyone at your organization fails to utilize their full $19/39 allotment each month, that's wasting money, whereas with API credits it is 100% utilized.
I don't think they've thought through the implications of this. Everyone should cancel and go usage-based billing with caps.
They do address this in the doc, Orgs can now (although it was vague as to whether it was an option or just the new standard, probably option due to business contracts) 'pool' the Usage billing across all users.
I'm guessing they did that (and the 'temporary bonus credits') to make the pill easier to swallow for that side of customers.
You're right, I missed that.
It still does make one wonder, why have seats at all though? If everyone is just in one big API credit pool - what do the seats/users accomplish?
It forces you to pay at least $20 in tokens per user even for people who use less (they probably have stats on how many people use just autocomplete, which doesn’t count against the quota. or have a seat and don’t use the service at all).
You can get at least at baseline vague stats but from what I have seen it is more account than user focused, i.e. 'X number of users used %feature%' and/or '%model% was X percent of requests per day'.
That said it's worth noting, I don't see how anything they expose will reliably help orgs plan costing from what AFAIK is in fact a big shift for billing/costing planning.
> It forces you to pay at least $20
For better or worse that's public pricing, i.e. if you are coming in also negotiating VS for devs, windows/office licenses for the rest of the business and stuff like Azure Devops... a lot of their stuff gets cheaper if your company's IT procurement group is vaguely competent at negotiating. Not even talking bigcorp here I'm talking 500-1000 employee range.
Of course, very small orgs will suffer, but it does tie in with the theme over the last two weeks; anyone with a personal account is basically subsidizing the credits for the business accounts during the transition period.
For orgs, each user was allotted their own quota. For messages beyond that quota, a pooled budget is available.
They mention in the announcement that it will be possible to pool usage across an organization.
This was my first thought too. "Oh cool, I should be seeing lower prices" as I don't use Co-pilot that often anymore. But no, that's not the case. It rather served to remind me that I should probably just cancel.
They could add rollover balances and be back to cell phone plans in the early 2000s.
Only if we also get unlimited nights & weekends.
$39 of credits at API costs is useless too, what are you going to do there, a single hour of coding? One half of a feature per month?
That's enough for 4 days of programming if you use GLM5.
Are you thinking something like rollover plans?
Everybody who says it's a 5-9-27x seems to not be aware of the obvious loophole. More like 50x increase. You were able to use over $500 worth of Opus on a $10/mo Github plan easily, no hacks. You could just prompt "plan this out for me, don't stop until fully planned, don't ask any questions", and you would get ~$5 worth of planning in one 3x request. At 100 requests/mo, each easily reaching $5, that's easy $500 worth of tokens.
Bingo. I created a few autonomous skills that did exactly that for plan review, implementation, and branch review, review autonomously until green.
I was using 100M+ tokens per day, $250 per day or so and only paying $160 per month to GitHub.
I cancelled my GHCP sub and switched to Codex last week, so far so good but I miss Gemini 3.1 Pro for UI work.
And this, right here, is why none of us can have nice, cheap things.
It was going to happen regardless due to the nature of enshittification. If they really wanted to stop people using 100M tokens a day, they could've prevented it years ago.
So, silicon valley decides to use their playbook of expand at all costs by burning money to acquire the market (like a carcinoma), and it is the users fault ?
Should we be blamed about uber destroying the taxi business, or airbnb the hotel one? Oh sorry, "disrupting".
Uber was dirt cheap, now it is the same price as taxis, and the people working for it (the "partners", not employees) have no social benefits.
Airbnb was cheap and humane, now it is THE cause for housing crises and massive residential property "investment".
The playbook of silicon valley is destructive, not disruptive.
It is by design aimed towards wealth accumulation. The ones with most money can capture the market, and make even more. It really is late stage capitalism.
And the more wealth inequality there is, the more pain, poverty and instability will be as well. AI will only exacerbate this.
Uber and Airbnb are not autonomous robots.
If people wouldn't use their services, nothing would happen. They would just go bankrupt.
So yeah, I'd say it's entirely people's fault. Because people just wanted to use their services without thinking what they're causing.
Customers who think only about themselves and noone else.
> Customers who think only about themselves and noone else.
When was this ever different? And do you expect it to ever change?
I disagree completely. You cannot expect every consumer to be fully educated and aware of the consequences of their purchasing power.
This is the role of legislation, educated experts creating policies so that you don't have to do business analysis before making a purchase.
Would I pay 10x the price for tokens and be outcompeted by other companies, hoping that openAI will go out of business ? This is entirely unrealistic.
Yeah it was crazy. Nowadays I use pi with OpenAI GPT 5.4/5.5, which to me seems both better and more generous than Claude. I supplement it with OpenCode Zen to get access to a bunch of models at token cost, and OpenCode Go ($10/mo) to get subscription-style access to Kimi, GLM and friends.
What is pi?
The minimalistic harness famous for having a tiny system prompt (avoids context pollution), and being what underlies openclaw. (https://pi.dev)
That was not my experience. When I tried to use Opus for longer tasks with Copilot, it would fill up the context completely and then crash without any output, while still consuming premium requests. (At least from September 2025 to January this year. Haven't tried after that.)
Copilot has improved immensely in 2026. I'd say to give it a try again if you're up for it. It works about as well as Claude Code these days in my experience.
So enormously that the haven't had any sane permission system built in this year.
Unfortunately, Opus was removed from the student plan in March. So far, I had been happy with GPT-5.3-Codex, but that model seems to have been removed this morning.
On pi coding agent, it worked very well for me over the past few months, but started glitching more recently, just prior to this announcement.
This might be Pi being buggy as heck.
When using opencode or copilot CLI, the error messages are displayed normally and it's possible to see what's going on. Under Pi, it sometimes just hangs, or Pi crashes with some bun stacktrace and that's it.
Copilot has introduced additional limits for Claude models in past month, and it's rather easy to hit it. Pi often doesn't show anything when this limit hits (although sometimes it shows the error, I guess it depends on Pi version).
Even more so, questions and user answers from agents were not charged as separate requests.
And when you make your harness ask you for next steps in a tool call, the journey continues forever, yeehaa
Yeah, people learned.
I created a 4 subagents that polled for new tasks, and restart after ~5h.
It was a great run.
This was my solution to very very though compiler tests that would take sometime up to 4 hours to figure out. Some of the time would be spent on running the tests, but still... I was burning so much tokens. I have free Copilot for my open source work so I wasn't even paying the $20.
this is the project that I am working on https://github.com/mohsen1/tsz
I did many 1h+ sessions of agent asking questions, delegating to subagents - all for 1 premium request.
I would say its a x1000 increase in price for agentic workflows.
exactly why i loved Github Copilot, you could pull of these shenanigans, and nothing would ever happen. That was the best part of it
but now, you get literally nothing
The 27x multiplier on Opus is the tell. That is not a pricing model designed for broad adoption, it's a price signal that says 'use the cheaper model.' The problem is that once users start self-censoring which model they reach for based on cost anxiety, you've degraded the product experience in a way that's invisible in the metrics but very visible in churn.
Flat subscriptions had one big advantage: zero cognitive overhead per request. That's worth more than people admit.
"Plan prices aren’t changing.”
Isn't this like saying "The Porsche you rented at $200/mo is now a Honda. But the price hasn't changed!"
This may be a more accurate analogy... "The Porsche you rented at $200/mo now only allows you a maximum of 100km of travel. You will be automatically charged extra when you go over that."
A whopping 100km per month for the low price of $199.99!*
* with a quota of 138 meters per hour, overage charges may apply
more like 100m
On top of being worth less, the subscriber discounts are gone.
The old plans were $0.033/request for Pro, $0.026/request for Pro+ and $0.04/request for pay-as-you-go. That discount is now gone. They even still advertise "5x the number of requests" for Pro+ over Pro.
Having rented cars a lot, this is actually quite common.
What country, I haven't seen anything but unlimited miles for over a decade.
Yeah, if I go to a petrol station with 50€, but only get a tenth of the amount of petrol I got last week, I may think that the price has in fact changed.
"Your monthly fee isn't changing but it now only covers about 3 days of driving."
More like, The Porsche you had for a month you'll now have for 5 min only.
More like, the rising gas prices aren’t a problem, I only ever fill up for $40
It's more like saying, "and you may now only use the Porsche for 5 minutes out of every day."
Full brake on the autobahn if you hit your 5min limit
It’s technically true that the plan prices haven’t changed, it’s just the value you get from those plans has plummeted. It’s classic deceptive sales language.
They are now charging per gallon instead of a flat rate per trip
I was curious why a company would still use the VS Code + Copilot sidebar method for coding, rather than something like Claude Code. Turns out there’s a GitHub Copilot CLI!
I thought I was pretty familiar with available options, but no one in my circles ever mentions this product. It doesn’t seem to have much mindshare.
Has anyone used it? What’s your experience?
https://github.com/features/copilot/cli
I'm curious about the opposite: Why would anyone use the CLI when, at least with Copilot, the VSCode plugin is super tightly integrated with VSCode, meaning the agent can see everything I can see. There's no mismatch in linter calls where I can see a lint in the ide that the agent can't find for example. I've had this problem even using CC in their VSCode extension, so I can't imagine it's not an issue in the CLI as well.
What's actually better in the CLI?
We need sandboxing for any agent, so we run it within Docker - so we use CLI.
I use vscode with containers extensively. Not sure why containers imply CLI.
do you run vscode in the container ? if so, can you share your config ?
i've been trying to do this with systemd-nspawn
Easiest way to do it would be enable ssh in container and then use VSCode ssh extension. Your host VSCode “becomes” your container VSCode.
I run it natively on my rocky8 container with UI sometimes but usually just do ssh
I use the Claude Code VSCode plugin for 80% of my work.
I prefer it because I can look at the code (although not as often anymore) and config (very often!) easily.
It also lets me jump to previous conversations easily.
There are a few cases where the CLI makes sense. One big one is if you are running multiple simultaneous sessions on a remote server using Tmux to have them preconfigured when you reconnect is nice.
Bun in general I don't see the benefit either.
You can look at the code in editor or IDE even when CLI agent is doing work.
I do that when I want to, but for me using agents in IDE is like looking with one eye covered.
The vs code integration is pretty slick. I can copy and paste function names into the prompt and it automatically turns them into these `#sym:` reference objects that I presume populate the context window with metadata about the function and where it lives. It knows what file I'm currently looking at as I jump around in the code, and that automatically gets loaded into the context. I can also drag and drop folders or specific files for context into the sidebar.
It's a lot of stuff that makes me have to type less into the prompt, since it's already getting so much info from my editor
I’m actually trying to move back from the Claude Code style, I feel like it’s easy to become distant from your own code, and I am feeling uncomfortable with that.
I’ve “vibe-coded” some projects and when I start to find issues or go to refactor them I don’t have that memory of why decisions were made, because many decisions were never made.
I've used it quite a bit. There are a lot of AI terminal coding products and this is another one. It works well, handles sub-agents without issue and does a reasonable job operating in the Copilot ecosystem. It handles mid-task questions and such we well.
I’ve tried OpenCode, Claude Code and Codex CLI. But was just shocked that Microsoft has a version I hadn’t even heard of.
Personally I got CLI fatigue and am happy with Conductor for now, but things are moving fast in this space.
The naming is bad. VS Code Copilot Chat.
But its a really good UI for agentic coding. Not sure why more people don't use it. I've tried the others and keep coming back to Copilot chat. It's a really good tool. Which is why the rugpull on pricing is so concerning.
Yeah, I've been using it heavily at work since the beginning of January (and have a personal Anthropic sub to compare to). Copilot CLI is pretty good, honestly. Most new features in Claude Code get cloned by Copilot CLI within a couple weeks. Claude models seem mildly more clumsy in that harness than the one they're trained on - subjective guess around 20% more turns for an equivalent task - but it's not a noticeable difference in the final output.
The other cool thing is Copilot SDK, so you can build agentic capabilities into apps, or build tools, that leverage the agent harness of the Copilot CLI:
https://github.com/github/copilot-sdk/
> I was curious why a company would still use the VS Code + Copilot sidebar method for coding, rather than something like Claude Code.
I use Claude Code, but I kept my Copilot subscription around mostly for really cheap usage of other models when I need to try a different one (which appears to be ending, in a sense) and also the autocomplete in Visual Studio Code which was really great across a bunch of files, I could make changes in one file and then just tab through some others.
I wonder what other good autocomplete is out there.
> also the autocomplete in Visual Studio Code which was really great across a bunch of files <...> I wonder what other good autocomplete is out there.
I am in the same boat. I tried looking for tab/auto-complete implementations ~ a year ago and it was pretty disappointing. If that has changed, would love to know!
Windsurf has free unlimited tab complete, I use it as an IDE alongside Claude Code and it works pretty well. I think Google's Antigravity also has free tab complete but no idea if its any good.
Cursor (paid) and antigravity (free) come to mind.
That's espoused as the big reason for the price increase: most Copilot subscribed developers it seems have moved to "agentic usage" with the CLI and Cloud-based agents.
Which feels a bit like a kick in the pants for me as a developer that was primarily using Copilot for VS Code ghost text and very rarely used the Chat sidebar much less "agentic" tools.
Copilot Pro sort of made sense for my personal account when amortized across a year, but I don't want to "waste" $10/month on credits I won't use most months.
It's in their ToS to allow using Copilot subscription with OpenCode - https://github.blog/changelog/2026-01-16-github-copilot-now-...
Absolutely the cheapest way to get a lot of tokens through a solid harness for $10/month. Until now
Liability. I work in a tightly regulated market and I cannot use any service due to be in touch with PII data. If it was my option I would just use OpenCode with deepseek, it is more than enough for the majority of my tasks.
I've used it. It's on par with OpenCode imho.
I'm just so confused why people aren't just using ghostty/kitty/terminal.app and claude code. Compared to the other approaches I've tried, it's by far the most effective way to get performance from opus 4.6/4.7
I don't know about others, but I use Copilot more often than other apps because of its tight integration with the VS Code itself where I still spend most of my time working on other things while letting AI do some task that I decided to delegate to it.
claude/gemini/crush inside of agent-of-empires inside of ghostty in one window, and zed in the other for touch-ups.
I tried the VS Code + Copilot sidebar approach a few months ago. It was definitely rough around the edges compared to Cursor/Claude. In our corporate environment, we weren't even able to use frontier models.
Quite honestly I love the GitHub Copilot CLI. I pair it with Squad and it’s awesome.
https://bradygaster.github.io/squad/
If I have the same repo also open in VSCode, it’s also aware of that fact, so you can give it context (a file or selected lines of code).
I've been using it for a few months because Copilot was the only AI blessed by our corporate overlords. It's not bad, I would say it's about 80% as capable as Claude Code, which I've used extensively on personal projects. However CC was recently approved, and I'm betting that with these changes to Copilot pricing we'll end up dropping it like a hot potato.
Because Copilot is the only thing allowed at our corp
it's my favourite harness so far
Search has become so bad that I also struggled to find Claude Code alternative and made my own tight (not editors, not plugins, not agents, strictly similar to Claude Code CLI) list: https://github.com/omarabid/cli-llm-coding
The list is not long but there are quite a few options. Even Grok has its own CLI!
The reality is, even though a CLI prompt looks very simple, it's a very complex piece of software. I personally use Claude Code (with GLM) and anything else I have tried was significantly inferior (with the exception of opencode).
I don't use Copilot or any paid AI but all of this usage-based billing reminds me of cellphones back when you paid per individual text message.
Usage paying for AI is 1000x crazier because you're not even getting a guarantee in the thing you pay for in the end. You have to keep feeding it prompts and hope it gives you the solution you want. You may end up with no expected result yet you are paying for it. At least with texting, you got what you paid for.
I wonder how long it'll be before all AI costs are flat unlimited monthly fees or even free across the board, without compromise.
I expect in the future we'll find out that someone in the industry was juicing the numbers with fake thinking tokens or something. The whole pricing model of charging you for the tokens it generates while not knowing how much it is going to generate going in has always been pretty crazy.
It reminds of early smart phones when the cell providers pulled away from unlimited data...and then they brought it back in s few years.
I think competition will get fierce. We see many people are attracted to the price stability of GHCP - it became clear what a request could do - the problem is that they didn't match results with cost. It's not clear what a 5 hour usage window in Claude Code can do.
There's no reason the harness couldn't provide a quote on the next request, aside from it takes effort and it would be upfront to the user, creating expectations.
Yeah, this was my frustration with Suno and Sora. You can burn a lot of credits (not to mention time) generating things that aren't what you wanted.
I don't mind a PAYG model for a simple chat interface. But when it comes to actually producing things, you burn through TONS of tokens creating the wrong output.
Internet usage when you billed by the hour but your connection was so slow it took a minute plus to load pages lol.
It incentivizes you to do most of that prompting on your own hardware/time, and only feed the final prompt with only necessary context to the big AI in the sky. It might even force you to think about the problems yourself for a bit!
> I wonder how long it'll be before all AI costs are flat unlimited monthly fees or even free across the board, without compromise.
That's already the case if you can self-host an LLM; you don't even need a mythical H200: gamer-grade GeForce cards can get you a long way there (if this page is to be believed: https://www.runpod.io/gpu-compare/rtx-5090-vs-h200 )
...after RAM prices return to normalcy, of course - and then wait another 2 or 3 generations of GPU development for a 96GB HBM card to hit the streets - and also assuming SotA or cloud-only LLMs don't experience lifestyle-inflation, but I assume they must, because OpenAI/Anthropic/Etc's business-model depends on people paying them to access them, so it's in their interests to make it as difficult as possible to run them locally.
Give it 5 years from now and reassess.
That page compares models that easily fit inside the ram on either GPU. The biggest difference comes when one card can fit a model and the other cannot.
Windsurf made a similar change in March: https://docs.windsurf.com/windsurf/accounts/quota
> In March 2026, Windsurf replaced the credit-based system with a quota-based usage system. Instead of buying and spending credits, your plan now includes a daily and weekly usage allowance that refreshes automatically.
With hindsight, per-request pricing makes no sense at all if an agent can burn a widely varying amount of tokens satisfying that request. These pricing plans were designed before coding agents changed the dynamics of token usage.
I wouldn't call it hindsight - I don't think anyone, at any stage, thought running a 10 minute+ sonnet session for 1 premium credit was ever profitable. We all knew it was a loss leader to get people using it.
It would have been profitable if that premium credit cost more than a negotiated discounted rate with Anthropic. We have no way of knowing if there were negotiated rates though!
There is no way to make that cost model profitable consistently. If 1 prompt can mean 100's/1000's of requests over hours, and you only pay for that 1 premium prompt, that can never be profitable.
They can engineer the harness to limit the amount it does. When pressing enter, it's be nice to have a "budget" per prompt, much like the model multiplier. When the harness used up the budget, it cleans up and cuts off the work.
But that would entail actual work and effort...and care for user's time and money.
Guys, you're discussing a house of cards to begin with: No matter how you're paying for the $CURRENTSOTA you're not garunteed that next month what you pay for will be the same.
So, lets do some honest evaluations:
1. The model itself is a non-deterministic engine of work with an unknown value; it's real value is just magic.
2. The business model itself is non-deterministic engine of profit with a known value; whatever the VCs have put into it, _must_ be piulled out. If Ed Zitron's numbers are correct, circa 2030, it's several trillion dollars.
So do some matrix multiplication of non-determinism vs determinism, and realize that the value proposition for _you_ is only going to decrease because #1 can never outpace #2, ensuring enshittification captures a smaller and smaller whale.
We know this. This has been the last 2 decades of money extraction from software. It was ok when it was some 12 year old's parents CC. But now it's you, or your business, that's going to either ben squeeze for value or squeeze out of the market.
And everyones squabbling about the color of the cost. ok
The problem with assuming that tokens can only get more expensive is that the Chinese open weight LLM firms have dropped models which have a known, fixed price that can never get more expensive (since we can run them on hardware we own).
Well, I guess we're not discussing the same thing. The cost of cloud tokens are going to go up. They won't ever be cheaper. They're generating far more tokens than my AMD 395+ w/128GB at a much cheaper rate.
I agree though, it can't get cheaper than the cost of hardware it's just without sufficient documentation of the actual costs to run the cloud models, we can't really know what the "true" cost of each token is. I assume there's an economist out there somewhere that could figure it out though. Certainly, the cost should approach at a minimum a open weights model running on a local machine.
I've succesffully got Qwen3-coder-next to loop and generate sufficiently competent code and from what I can tell, the difference between this and the cloth is how quickly the gen happens and perhas how interactive it has to be.
per-request was broken, yeah. but $10 of monthly credits is basically just a prepaid wallet with a reset timer.
I wonder if GitHub (Microsoft) is implicitly betting that enterprise demand is sticky enough to absorb these rates, especially given that Opus 4.6 “fast” was being listed at a 27x multiplier. Maybe they saw enough usage at that price point to conclude the demand is real. Or maybe the strategy is to keep the enterprise customers who can justify it while shedding heavier individual and power-user usage.
The interesting question is how long it takes enterprises to notice the capability/pricing tradeoff, and whether they respond by limiting access to the strongest models internally.
The part that worries me is that this market is still very early. Most developers and organizations are still learning how to use these tools effectively. Raising the experimentation cost this much may slow down the discovery process that makes the tools valuable in the first place.
As someone who works in a Microsoft shop with Copilot, I am curious to see what happens.
Due to data governance it will be difficult to move to a different provider.
At the same time, this price hike is so large that the ROI on copilot will be a net negative.
I think what will ultimately happen is that we will not pay Microsoft more than we currently do and we'll simply end up with less AI usage in the company and a reduction in productivity.
I also suspect that there are many "slow-moving", Microsoft heavy enterprises but with in-house devs that can't get anything but Copilot approved, and Microsoft trusts this will remain so.
It's not turning consumption based because there are a ton of these licenses just sitting idle.
As a single data point, this is absolutely true. At my current "Big Corp", Copilot was immediately approved while Claude is entering month 2 or 3 of trying to get approval.
Additionally, we got copilot for every user, including those that never write code or use AI tools.
yes. even in my province that's relatively behind on tech, at least +-20% were already strongly relying on Microsoft (Azure) and thought adopting Copilot was a no-brainer. That ofcourse comes with autonomous programming, because we're paying for it anyway!
These companies already lost the spark/innovation years ago. They're just using LLM's as a way to survive. Wonder how long that lasts.
Opus 4.6 "Fast" was originally at 10x, literally the same cost as Opus 4.1. After promotion period, it was 30x.
As someone that is on the enterprise side in a non-tech F500 company, what I'm seeing is some FOMO and need to be part of the hype cycle. We're about to plonk a bunch of money on more Copilot licenses. Something got in the water where all the C-levels the past two months are pushing everyone to use AI but when they bring up examples of their uses its like "I use it to rewrite my emails" or prompt 'engineering' ideas that point more to patching over poor processes, data management, and decision-making within the organization or not.
What we're seeing across the board is every software company tossing AI onto their name or sales pitch and no one understanding what that actually means. But we will spend money on it because of FOMO.
I really question if we're reaching the end of the hype cycle to the point. I wish I were brave enough to put money on it. It feels like there was a command from up top to 'do something with AI' and leadership is scambling for some resume-building projects vs doing the hard work they should've done the past two years at a people and process level.
I'm at an org, where every medium to large company meeting talks about AI in similar breathless non-specific terms, yet m365 copilot is the only approved AI chat, and GH copilot the only AI coding tool.
Subsidies stop when LLMs improvement plateaued (though they still benchmark higher somehow). At some point, you have to make money or at least break even; and I think they concluded that we reached that point.
I really hope that they think so and that they're wrong and they get burned hard. Them and all the AI labs that lied, stole, inflated, hoarded and tried to justify all this as an existential moment where AGI would radically change society. I hope their calculus to reel in paying users is all wrong and now they all crash and burn instead of recouping VC money.
Github had, by far, the most easily game-able agent usage policy. People would force the agent to run a script before the end of turns that consisted entirely of `input("prompt: ")` so that you could essentially talk endlessly to an agent for the price of a turn. I see this less about the future of this industry and more about fighting the costs incurred by bad actors.
I never played any games like that, but simply giving the agent a clear exit criteria and instructions to check the exit criteria every time it thinks it's done on a complex task was often enough to keep it chugging away for most of a day on a single prompt in my experience. Per-prompt pricing just isn't sustainable period, even if everyone is acting in good faith.
Charging by prompt was always wild to me.
I once asked it to do a comprehensive security review of our code. It churned for nearly an hour (and then produced 90% false positives). Insane that that usage was charged the same amount as me just saying "Hello".
After 2 months of using copilot pro, they've charged us 22$ in premium requests for my user, i spend roughly 1900$ in tokens, going of the on-demand pricing. This is estimated to be around 40-60% of real costs. They are undercharging by a factor 50 to 90!
This is just the start of the rug-pull
If the real cost where really 1900 / 50% = 3800 USD per month then you could rent your own H200 with 140 GB VRAM.
The issue currently is capacity. Servicing the models cost just peanuts of electricity but the popularity is killing them.
There is noticeable trend across all agentic coding platforms that this situation is no longer sustainable.
With this kind of pricing (sonnet 4.6 has 9x multiplier, previously 1x) it begs the question why use Copilot to begin with.
You could easily just buy the tokens directly and have a lot more choice as well.
One reason I used it was that I wasn't locked into a single provider and switching them was as easy as changing a drop-down. Small feature? Sonnet or GPT5.4/mini? Large changes? Opus. And why not see how good Raptor Mini does this one refactor?
It also helped build an intuition of what wach model could do and which parts it was weaker at because you could try them almost side by side, especially if one model's output wasn't great.
That said, these were all side projects so nothing truly consequential. Otoh, you might leave some extra perf on the table but I found the models worked quite with the Copilot harness.
Yeah, this is a very useful abstraction layer. The entire concept of separating the model creator from the model runner is good for competition and is customer friendly. Which means they likely hate the concept and want to kill it.
Gosh, imagine getting to do that with your TV/Streaming subscription. Getting to pay one fee to access some set number of hours per month from any of the providers.
The problem is I can't afford the tokens! Even on my $10/mo plan, running either 100 opus, or 300 sonnet agent runs would cost hundreds of dollars - well above my budget!
Doesn't GitHub get volume discounting they can pass on to their Copilot customers?
Economics of scale don't work when scale still isn't enough and capacity is still limited.
GitHub has the full power of Azure with their hosted models but it's not being passed to consumers.
Economics of scale don't scale
It seems to me more expensive but I might be reading it wrong.
Looking at their pricing it does not look the case.
I liked copilot because I didn't have to think about tokens. I get hung up when having to think about the price of things, and its hard to think about the project at the same time I got to think about token usage like a gas bill. The usage system had its own issues, but having a set amount of requests was a very comfortable way to use a paid AI service.
Sounds like you're a candidate for a local model. It's kinda nice not caring what the token count means except as to compaction.
Not paying per token? Not sending my code to someone else's servers for inference? That's the stuff of sweet dreams for a stingy, paranoid solopreneur like me.
If I could run a local model comparable to even Sonnet 4.6 without shelling out $50K in hardware, I'd do it in a heartbeat. But all I have is a 32 GB of RAM and an old RTX 4080.
Or am I not up to speed? Are there decent coding models that can run on dev laptops? Not that that's what you were suggesting by recommending a local model, necessarily; just curious.
I am trying to figure this out too... what I am seeing is that the local models like Qwen 3.5 family that fit on hardware like yours handle ambiguity poorly. But are capable of emitting complete apps too.
That, and they have tool use issues.... https://www.reddit.com/r/LocalLLM/comments/1smzw6s/qwen35_a3...
I would check out the model mentioned in that thread, GGUF unsloth/qwen3.5-35b-a3b on Q4_K_M
Qwen 3.6 is out now and a touch better than 3.5.
I'm finding Google's Gemma 4 even better though - seems to hold up the agentic loop better than Qwen.
All will load into 20Gb of VRAM. None are amazing, but they do just about work.
I do love using local models when I can, but qwen-35B is the best model I can run, and while its an insanely good local model, it does not compare to the big ones.
Have you tried the latest Gemma? You might prefer it to Qwen, depending on what you're doing.
I did, but in almost everything I tried even qwen3.5 was better, and 3.6 was a huge step up.
Good to know, did you also try non-coding or is that what you use it for? For tasks unrelated to coding we've generally preferred Gemma.
So far I have done coding projects only, but I have some ideas for some non-coding agent projects in the future, so I'll give Gemma another shot
So given that I primarily interact with LLM's through VSCode, and I prefer the Copilot interface to the Claude Code plugin, does anyone have any suggestions on other plugins I should try? In my experience, Copilot is much more "plugged in" than any of the other plugins, in the sense that it can see things like linter outputs in VSCode. Basically, copilot "sees what I see" in a way that no other plugin or command line tool can, which make it much more ergonomic to use.
With this pricing change, I see no reason at all to stick with Copilot in principle, but I really need to solve this issue of IDE integration to move on.
You can use Copilot Chat* with basically any API provider, and if you switch to the VS Code Insiders build you can configure it to use literally any OpenAI API-compatible endpoint.
Other than that Zed has a similar experience which is pretty decent.
* By which I mean the good one, whatever it's called now - the part of Copilot that used to be a plugin and is now part of VS Code, not the thing that has always been part of VS Code.
Try Cline, a vscode plugin that's been around since the start and has really active development. You can also use most providers as well.
Saw this new Cline fork mentioned in another post: https://github.com/dirac-run/dirac
Also heard of more and more people moving to Kilo Code or OpenChamber instead.
I tried Openchamber. Far too buggy. Cline seems like the best bet.
I used to feel exactly like you do, but now I use the Claude CLI exclusively and am very pleased with the results.
Could you elaborate on what's better about the CLI?
Simply the output of the code written. It initially worried me that it was harder to add context (can't select a single line of code in VS Code and automatically attach it, for example), but I've found that the Claude CLI's output is so far superior that it doesn't even need that VS Code context.
Are you telling me that inference costs did not go down, as the AI crowd keeps preaching?
Cancelling. Going with Codex $100, Kimi annual plan, DeepSeek API, and a local LLM once I get a Mac Studio.
Inference economics are going to be brutal in 2026 H2 when DeepSeek's new infra and model improvements come online, and Kimi launches K3. By brutal, I mean for OpenAI and Anthropic.
What about Minimax? Their api pricing is half of Kimi and Z.ai.
Their models are too small for coding. Maybe with M3, we'll see.
Has anyone found the answer to this yet?
> What is the benefit of using the Copilot Pro+ at 39$/month instead of using the Copilot Pro at 10$/month and paying for extra usage?
Some models, for example Opus 4.7 and GPT 5.5, are only available on Pro+; Pro+ has audit logs and GitHub Spark; that's about it, as far as I can tell from https://docs.github.com/en/enterprise-cloud@latest/copilot/g...
Yes there's no Opus at all on Pro. GPT 5.5 is also missing. Then again what would you expect, the economic reality is beginning to hit. Also I can't be too mad when the "base" models (GPT 5.4...) are still available and decent.
When I see how fast Codex max thinking GPT 5.5 eats our enterprise seat credits almost anything else seems cheap (until we switch our live systems from 5.4 api to 5.5 api I guess)... good thing I'm not the one paying for those credits and tokens (which is probably how most of the money is going to be made on AI going forward, borderline free chatbots for normies are done)
I'm using Sonnet 4.6 all the time and it fits 100% cases for me. People overestimate Opus, as well as GPT 5.5.
5.3-Codex is really good enough, Sonnet 4.6 is good enough.
If I had to guess...
On my personal account, Copilot Pro+ still only gave me back Opus 4.7, whereas my work's Pro account still lets me use Opus 4.6.
So, my gut says, it's entirely possible that Pro+ will continue to have more segregation on model availability...
FTA
> Last week, we also rolled out temporary changes to Copilot Individual plans, including Free, Pro, Pro+, and Student, and paused self-serve Copilot Business plan purchases. These were reliability and performance measures as we prepare for the broader transition to usage-based billing. We will loosen usage limits once usage-based billing is in effect.
There's enough weasel wording here that I would expect only certain models get re-enabled on Pro.
e.x. lots of people seem to get good enough results from Opus 4.6, personally I prefer it over 4.7 in GH Copilot... locking that down to Pro+ would be, given this salvo of enshittification, a 'logical' move on their part.
vindication! https://www.wheresyoured.at/exclusive-microsoft-moving-all-g...
You should indeed feel vindicated. Me too, I've been a paid subscriber for months and eagerly read every word. <wink> But yes, I feel like this is the first of several vindications you have coming.
I thought of this as soon as I saw the headline: I hope this is the start. Crossing my fingers to see more line up with your predictions.
What's the current situation for coding with Local LLM's on decent hardware? I have an M3 Max with 64 gb of ram and am thinking I should start looking at Ollama and Opencode? Is this a useful stack for smaller personal projects?
It’s getting there. You could give a try with qwen 3.6. It’s worth paying for better models in the cloud, but local models are now better than nothing.
One nice development recently was ollama's support for MLX optimization on Mac hardware. It's not obvious how to know you're using a model that works with it, yet, so it's rough around the edges.
https://ollama.com/blog/mlx
Use llama.cpp or better yet Unsloth Studio
After seeing the ridicolous multiplier increase I've added a calendar event to cancel my subscription mid-May.
(I'm a copilot subscriber since 2022)
Already cancelled here! One less thing tied to GitHub.
Current multipliers vs from June
EDIT: only applies to annual plansI think that only applies to held-over users on the annual plan:
> Users on annual Pro or Pro+ plans will remain on their existing plan with premium request-based pricing until their plan expires, however, model multipliers will increase on June 1 (see table).
There will surely be corresponding, different AI credit costs for each model.
"AI credits" are just US cents. The cost per model is here - and it looks like it is just the API providers API cost: https://docs.github.com/en/copilot/reference/copilot-billing...
Kind of. Monthly users moving to an entire different pricing model so we don't really know what the increase in price will be for them.
Thanks this is helpful. I am on my first month of annual Pro. Should I keep it or cancel it (ask for a refund)?
It isn't just the big multiplier increase, they also say "...and no new models or features will be added to annual plans going forward."
Can you imagine ten months from now and you're still rolling Sonnet 4.6?
Cancel/refund is looking pretty good. They're doing refunds until May 20.
"To request a refund, go to Settings → Billing and licensing → Licensing, select Manage subscription, then choose Cancel and refund "subscription". (The phrasing varies slightly depending on your subscription ). This option will be available until May 20."
Appreciate the info. Have you got a similar opinion of the 'convert to monthly' option they've also provided?
Not apples for apples.
Before:
- Opus 4.6 each premium request is 3 premium requests
After:
- Opus 4.6 each dollar spent is 27 dollars in copilot AI Credits.
Given that you'll receive 19 dollars of AI Credits in Business plan, that means you can probably say 1 "hi" to opus per month.
It is an apples for apples comparison since those new multipliers only count if you are on an annual plan in which case the premium request system stays in place until you either cancel and get a refund or until your renewal comes up. https://docs.github.com/en/copilot/concepts/billing/usage-ba...
If you are not on an annual plan, multipliers will be gone completely. You can see the rates that apply instead here: https://docs.github.com/en/copilot/reference/copilot-billing...
Based on the pricing and comparing to competitors e.g. bedrock[1] looks like cache-write will only be on 5 minute TTL.
[1]: https://aws.amazon.com/bedrock/pricing/
Thanks for that link. Oddly the blog post didn't contain any of this information.
GPT 5.4mini is even worse, from 0.33x to 6x which is ~18 times more expensive now.
GPT-5.4 and GPT-5.4 mini are now the same price, which, why?
Those multipliers will only apply if you are currently on an annual subscription (and only until your renewal comes up or you cancel). So I assume they simply want to make it as unattractive as possible to get most people to cancel it and move to the token based system.
That's not an answer. It's specifically a discrepancy between 5.4 and 5.4-mini. If you look at all other models/generations you see that the cheaper model indeed has a lower multiplier. It's very strange that only 5.4 doesn't have this.
both 5.4 were best bang (-mini even more as I found it usually same well performing!!!) for the buck before. Now they face cost reality it seems.
GPT 4.1 had a multiplier of 0.
I was surprised to find that this sentence
> Plan prices aren’t changing
did not continue with an em-dash followed by something profound that is changing.
Plan prices aren't changing -- the value you get out of it is.
They're not the only ones in the AI sphere to wind back, but they're the weirdest case in my eyes. Microsoft invests in having engineers building open models and they don't use a single one. I really don't get it.
But what really surprised me most about Copilot is that it would bill you per question, nothing about tokens. So if I managed to produce a prompt that gave me back an insane amount of tokens for something, which using any Claude model would easily accomplish, you were giving me my money's worth, at your own expense. The math is not gonna math out forever.
> I really don't get it.
One of the largest employers publicly engaging in a project which has the outcome of depressing wages. It's easier to "get" if you don't take the trillion dollar gorilla at face value.
The cheapest copilot plan felt totally unsustainable to me. For around £8 month i was getting 100 opus 4.6 prompts (albeit with a reduced context window size around 128k iirc vs 200k to 1m for first party hosted opus). Gpt5.4 was hosted with 400k context iirc.
On top of that, you’ve got 2000minutes of container runtime, so running cloud agents was included. As was anthropic agent sdk mode via copilot which is very comparable with claude code - not identical, the anthropic “modular prompt” is much leaner in the sdk version.
I cant say im mad, i got above what i paid in value. That said, going forward ill probably go back to openrouter payg rather than a subscription.
I got a free 3months of the gemini £19 plan and ive been playing quite a bit, 3.1 pro is a good model, i just find it slow. Flash i think i under appreciated until now.
I AM mad, because I just signed up for my private copilot sub. Otoh I do have it at work. I suspect we will start noticing differences there, too, but it all depends on the top users, vs average user since it is now pooled, IF they dont just allow extra usage. Bu that has become more expensive now, so not sure what they will do. I am hopeful that very few actually even use it. In my and surrounding teams, only maybe 1 out of 3 people really use it much.
I pay for Copilot annually, and mostly for its code auto completion features. I use CC if I want to do anything agentic. Not sure if I want to pay more for occasionally-good-intellisense at this point.
Same! I wonder what other alternatives there might be for autocomplete.
autocomplete is still unlimited within the subscription, maybe using a free model or even cheaper ones are best value..
If you find out, please let me know!
Code completions and Next Edit Suggestions remain free with Copilot Pro.
But you can no longer amortize annually, which makes it even more a question of "is this worth it this month?" each month. Especially for personal accounts.
I'm similarly thinking about sticking with the auto-downgrade back to Copilot Free when the annual sub ends and then just yelling about it any months I hit the 2000 completion cap.
I wouldn't mind a plan between Free and Pro that is just "all I care about is code completion and next edit suggestions".
This marks the beginning of the end of the AI free money era. Next they will dramatically raise prices when they have to be profitable on tokens.
I am seriously considering binge buying local AI inference hardware. The way this is going, there will be another big GPU crunch soon because everyone will need local models and/or open model inference capacity to do their programming tasks when the subsidized subscriptions are no longer flowing.
I just bought a 3090 off of Amazon. It was expensive ($1500ish), but covered under prime and an included any reason 90 return policy.
But that gives me a good while to determine if it's worth it or not. I've heard good and bad, so here's hoping for good or close to it.
I wasn't going to fork out $1000 on a chance it might be enough with a rough return strategy.
How is this legal when people paid for a yearly plan in advance?
In order to most-to-least charitable, any of:
1. Github could choose to grandfather in those plans and make no changes until those plans expire.
2. Github could offer, or the user could request, a pro-rated refund along with cancellation of the account.
3. Tough luck, those users agreed that Github could unilaterally change the ToS at any time.
Right now there is a cancel and refund button on the Github Copilot annual subscription setting, which I have just pressed.
> 1. Github could choose to grandfather in those plans and make no changes until those plans expire.
They explicitly stated that they won't be doing that: the multipliers go into effect in June for everyone, annual plan or not.
I doubt you can force them to provide the service with the original terms, but you might be able to ask for a (partial) refund. If not today, after a week of verbal abuse they will receive for this online.
It depends where you’re located. In the EU they have to honor the contract you entered, but presumably there is a clause that they can prematurely terminate the contract without cause and give you all of your money back (from the start of the contract).
In places with reasonable consumer protections (Australia, Germany) it almost certainly is illegal unless they give a full (whole year) refund. I think the short time limit of applying for a refund won't be looked at favorably either. Regardless of their ToS which I'm sure covers this.
But companies do lots of illegal things, and in general nobody takes them to court over it.
My thought exactly! First the usage limits + model limitations and now fundamental change to the billing. Hope some consumer watchdogs are looking into this!
100% it says in their terms that they can change the service during the agreement.
That kind of clause would be void in many places around the world.
For example, the German Civil Code states:
Let’s see if German users can enforce it or not.
I just checked and you can cancel with a refund.
For the yearly plan they only change the model multiplier. And it's in the subscription contract they can change that multiplier at any time.
So it may still be cheaper for us to train junior engineers in the long run.
It begins.
"It" being the end of subsidization of tokens and plans (expected) but while lock-in to foundational models and cloud services is still lacking. Guess investors want their ROI sooner than later, given how big of a wrench the AI boom has thrown into global economics.
It's almost impossible to afford the usage-based pricing as a heavy user with multiple max and pro subscriptions. Seems like a bet that same-capability inference will keep getting cheaper on something like a Moore's law curve.
“Base plan pricing isn’t changing” is technically true, but for anyone using the more capable models heavily, this is still a price increase in all the ways that matter. The old abstraction was hiding compute costs; the new one mostly stops pretending.
“base price isn’t changing”
Why does everyone care about gas prices I only ever pay $20 for gas?
Shrinkflation
I bought a copilot subscription for some small personal projects at Christmas.
I haven't been able to use my subscription much over the busy spring months, but i'm being charged every month.
I'd be tempted to keep the subscription if usage-based billing meant that i'd save money when i had less time.
But today, after hearing this, i cancelled my subscription.
Anyone paying enterprise billing can you shed some light how on the earth you will be able to fund this bill going forward?
Were you able to see assisted AI coding savings proportional to costs increase now you are going to get?
Companies removed people as AI assisted coding will be cheaper and now coding cost are going up from fixed $X to non-deterministic. The posts by Uber few days back about spending 12 months' worth of money in 4 months tells a lot.
Only path forward seems using Open-source models and many companies don't use Chinese that makes only Mistral one as the option.
Surely one path is to return to humans. We can accept that a profit grab to reduce staff numbers to maximise profits, maybe was never the best plan without being in control of the development tool.
Does this mean that "Let Opus 4.6 churn in a terminal window my whole workday for a fixed price my employer barely notices" was never sustainable?
Sample of 1 opinion: Massive downgrade for fun vibecoding on the go. Got done much before the recent rate limits on mobile on fun projects, their product went from fun to bad within a few weeks. And now into oblivion I suppose.
So I guess from now on GH Copilot is only worth it if you want a quality autocomplete in VSCode.
That was the first thing I turned off in VSCode. Autocomplete for my TypeScript projects was great. And the "AI" suggestions/completions were really getting in the way of me still being the "driver."
Does this mean you can only prompt "Hello" every morning for a month with Opus 4.7 ?
GitHub Copilot has been the most expensive LLM subscription on the internet, when it comes to dollars-per-request-limit. Literally any other subscription provides more usage per dollar. https://codeberg.org/mutablecc/calculate-ai-cost/src/branch/...
some of Github's open source maintainers have lost their free github copilot pro, guess this is really the next step for them to save cost in their infrastructure.
I started to use github copilot with vscode, but have never been too happy about the system. Over the months I gravitated to much more agentic workstyle, hardly ever editing much code by hand. The vscode IDE was getting more in the way. I had already started to look at OpenCode, and when I found it has a web interface, I was happy to switch over. I use a simple editor (KDE's Kate), or just less to skim through the code and/or a git diff. OpenCode has some free models in it, but I think I will need to get some kind of subscription for a better one. But it won't be copilot any more. The market is moving so fast that I don't know what are the most resonable models, or the most flexible way to set them up so I can switch when prices change yet again.
Check out Zed sometime, its pretty decent too.
Annual Pro+ subscriber here, mostly using it as a fallback when my Claude Max plan hits limits. The request-based pricing was genuinely the appeal, I could spill over from Max into Copilot without thinking about token costs. Going from 3x to 27x on Opus on annual plans is rough. The newer reasoning models with variable thinking budgets probably made per-request pricing untenable in the long run, but the way this lands on existing annual subscribers feels harsh. Going to look into the prorated refund.
I pay for github copilot solely for completions and use codex for actual agent work. I really wish I could pay like $5/month for just completions or there was a good local alternative for them.
What's the residual value of Copilot after the changes? For Enterprise plans, even copilot code reviews will be charged at token price + github action minutes for the execution of the review. One can roll one's own reviewer and have it spend the API tokens as well. At least one will be able to select the subset of files for the change set, if needed.
The background agents will also depreciate in value because of their harness that's a black box that's not optimized for token usage at all. Rolling one's own will be a better choice here.
Whose idea was this “premium request” model anyway? If you’re going to invent a new metric used to bill, why not align it with what, even at the time, was a clear underlying cost structure that GitHub actively chose to ignore for a more confusing system.
It made more sense in the ye old days where a request was basically just a chat message in a sidebar and it could also edit code. Then saying someone can use 300 chat messages a month kinda makes sense.
Turns out when a request can spawn tens of subagents and use millions of tokens over many turns of toolcalls then suddenly github copilot has a massive financial problem on their hands.
This approach started with the “Ask a question about your code” feature, which is more comparable to single chat message with relatively predictable token usage. Now it’s an agent who might work for 30 minutes, read the whole codebase, and write 1000 lines
30 minutes lol. I've gotten Opus to work for a full 8 hour day on one prompt.
I'm not usually a Conspiracy Guy, and the answer is probably `incompetence * tech_debt`. But I think that having sufficient layers of abstraction to any billing model is a useful way to hide the real cost of things. It's why it's done everywhere.
In business, if a strategy is making more money, the most direct and usually correct answer is malice, not incompetence.
> Code completions and Next Edit suggestions remain included in all plans and do not consume AI Credits.
This is the VSCode autocomplete stuff right? Really enjoy this.
> Copilot code review will also consume GitHub Actions minutes, in addition to GitHub AI Credits. These minutes are billed at the same per-minute rates as other GitHub Actions workflows.
That sucks.
Looks like Microsoft has run out of compute and can't scale it fast enough to serve copilot users and Azure AI Foundry needs, given that the customer base is growing there as well.
In light of this, does anyone have a DGX Spark and use it as a coding agent?
I'm happy I invested in local solutions and cutting context to the bone for API providers. Claims about AI being able to fully replace programmers never took into account the long-run equilibrium price of inference.
Already there are companies paying more for coding tokens than for programmer salaries.
That's my point. They made those decisions without any consideration for the long run, which would have required them to project the cost of AI services years into the future. Obviously management didn't do that and had no way to do that. It made current earnings look good, though, which was enough for them when they made the decision.
Not sure what your point is, but they are firing lower performing programmers so that the top programmers have more tokens to spend.
Doesn't necessarily mean that's a good idea.
I only use copilot for the occasional auto-complete suggestion. I'm betting I could run a lightweight local LLM with llama.cpp to get similar functionality. Maybe this would be a decent replacement https://github.com/TabbyML/tabby
I was always curious how they can sustain a request based pricing model when requests can range from tiny to huge with all the modalities GH Copilot offers. Was a steal for agentic coding that turned out to be too good to be true, in the end. Still: Thank you for the ride.
I guess this is the "Google App engine" of Vibe coding when google raised pricing of App Engine significantly after being in preview. Doing weird price changes like this when coming out for new services make much more sense than doing this change which will just anger users.
According to `bunx ccusage` I'm easily doing $250-400/day in "real" API costs on my $200/month plan. There's no way everybody else isn't going to do the same thing and completely change the industry again. Both beginner and advanced developers are already hooked on all this stuff and they all know it.
Built credit pricing into my SaaS for AI features and the hardest part wasn't the math, it was that customers can't easily predict their own usage. They underuse and feel cheated, or overuse and churn. Subscriptions hide that volatility from the customer. Usage based pricing makes it their problem, which is honest but harder to sell.
We all knew this from the very beginning but couldn't compete with OpenAI or Anthropic on their subscription-based pricing strategy. It was nuts except for those few corporations burning investor money to keep competition out as long as possible. Now they don't have to hide anymore that subscription pricing won't do it for ai. The pyramid scheme is falling.
Here goes my Copilot Pro subscription then, reluctantly heading over to Codex CLI since the CC base plan is downright unusable.
I'd be fine with it if they can make Sonnet 4.5 unlimited. I haven't personally seen any major differences between 4.6 or any of the other newer models. Claude 4.5 seems to be the right balance and works great for me.
Just got an email with this announcement.
I have Copilot Pro that I use occasionally, but not enough to tell how the switch to per use would affect my usage.
Based on description Pro plan users will get $10 in monthly AI Credits, but that seems rather low compared to what you could use same plan until now.
> rather low compared to what you could use same plan until now.
That's exactly where the subsidy is being removed.
The problem is that people expect to get the output of 100 people with a $20 subscription by spawning multiple agents. This is unrealistic. I'm using 2 codex plus account and able to manage a repo with 265-300k lines of code.
The point is - if its the same or more expensive per month than a real human employee - why pay for AI ?
Human retain knowledge, product knowledge, can pick up more work often for the same money. And having many of them means your business wont go down if provider suddenly bumps API pricing.
How much is the average pay for a junior developer in US? Its definitely much costlier considering PF and other benefits. Just the math. if you use it efficiently it's much cheaper than hiring a permanent staff. You can maintain a lean team and do all the mundane boilerplate coding with AI.
Time to cancel. Happy to hear about some alternatives.
As a Github Copilot user, who mostly just uses chat in the VS Code editor but still burns through my Pro limit every month -- what's the best alternative price to performance? Claude Code?
I keep hearing that Codex is the best bang for your buck now.
"Paid for annual? Tough luck, from now on your usage limits are reduced by 89%. You can do 11% of what you paid for. Good luck if you paid annual a month ago!"
And then they have the gall to say
> "The bottom line: Plan prices aren’t changing"
If anyone lives in a place like Germany or Australia and has an annual sub, please take them to court, you're guaranteed to win because you have reasonable consumer protections and their ToS doesn't stand a chance. 9x reduction is unreasonable and the consumer cannot be expected to see this coming.
In case some diehard enshittifier believes that consumers should know better and businesses should be allowed to get away with it, where is the line? 99% reduction? Is that still okay?
If this situation is to be acceptable then it should be regulated as a financial product like stocks, which come with knowledge tests of "do you know you can lose all of your money?". And come with regulatory compliance and all that.
In a way, this is good news. I am expecting to see a lot less of AI-boosters claiming that they are "productive in ways you don't understand", now that they'll have to dish out upwards of a few thousand bucks per month for whatever PoCs they are building. And the excited hired middle managers and CEOs spending other people's money will have to be much more careful about their budget starting June.
Nah, they are moving to DeepSeek v4 Pro which is like Sonnet tier for a fraction of the price.
And DeepSeek V4 Flash is comparable to Gemini Flash at also a fraction of the cost.
These models are just quite chatty from my experience. Only downside.
And DeepSeek is planning to deploy to their own Huawei TPUs, halving their already cheap prices.
This reminds me, I have to cancel co-pilot...
Soon it will be cheaper to just do it yourself
It may already be cheaper if you calculate the hidden cost accumulated by LLM generated tech debt. At least some of it could be avoided by manually doing it better to begin with.
I searched for "Show HN" for the first time ever and I got the impression that most of the costs associated with the projects could have been avoided by never starting them to begin with.
"VibeBrowser", "Financial Database API for Vibe Coders" followed by dozens of self feeding AI projects with barely anyone doing anything that isn't related to AI.
so what's everybody using to get autocompletions in vscode? i've been using copilot just because $10 is cheap, but i use opencode for everything other than completions.
i tried the continue vscode extension, and it seemed kind of janky. are there better options?
Windsurf has free tab complete and has been good for me. Antigravity has free tab complete as well but no idea if its any good
I tried antigravity, it worked great for a couple weeks and then Google changed their rate limits and ruined it - I’d get about 3hrs of work before all the ai features just stopped working.
from these comments I'm going to make mouse.dev pricing as follows
-BYOK runs are $0 to Mouse, period.
-Hosted runs are billed at provider cost + a published markup.
-We will never invent a unit of billing that isn't denominated in tokens, seconds, or tool calls.
-Credits in the paid category never expire.
Why would anyone stay on the Pro+ plan going forward? Pro with openrouter for Opus would be cheaper?
just cancelled mine. Only reason to have it was I could get a lot more done with a single prompt. No reason why I shouldn't go to the model provider directly instead.
It was just a matter of time, considering how many Mtok you could consume in just 300 prompts.
So what's the best alternative now? Openrouter + cline or something else?
i like opencode's zen https://opencode.ai/en/zen with opencode's TUI.
Well, the free launch for somebody else's money is over.
End of an era for predictable costs as a small business. We will refer to these times as ‘the good old days’.
I expect the prices to rise to almost cover the cost savings from layoffs at large companies.
Amp Free provides $10 free credits every day. Unfortunately, new applications have now been closed.
what's the best place to transition to on a company level after this change ? should companies just switch to pay as you go with open router ?
This subsidized inference is just a marketing ploy to increase prices and profit.
If common people can have a DIY setup with an open source model cheaper than those behemoths with a scale advantage, it's clear that we have been played.
Time to either self host a Chinese open source model or to just pay the cheap Chinese providers.
Yeah, local is clearly the future. Even beyond the cheap Chinese models you can install the apfel[1] stuff if you're on a mac and want a quick available onboard cli option. And I'm sure people will adapt the Flash-MoE[2] integration to be even better soon as well.
[1] https://apfel.franzai.com/ [2] https://github.com/danveloper/flash-moe
I really don't understand why OpenAI, Anthropic and Microsoft are in competition to see which one of the three will elevate deepseek the most.
DeepSeek will do the same thing.
Z/Mimo already raised their prices multiple times since the promotional prices at the start of the year.
DeepSeek (and other open weight models) can be served by anyone with the means to do it. All of the big open weight models are available on OpenRouter via commodity providers.
DeepSeek is actually more expensive than expected ATM due to compute shortage, but they said it would come down in price.
The compute shortage will be bigger 3 months from now.
Depends how China scales its semiconductor industry.
cursor, windsurf, and CC are all already on usage-based models so I guess what really matters is whether Copilot's GitHub integration depth justifies the price per token vs the alternatives
And so it begins...
Now do the same thing with the office 365 subscriptions.
That is, redact copilot from the bill for customers that dont use it.
haven't touched copilot for one year, this reminds me to cancel, it will take sometime to catch up
So about a month left before cancelling. Got it.
I'm not sure I understand this. All I know is now, I pay $39/month (actually less because I paid a year up front), use the agent, mostly on auto--and only choosing a model if it got stuck or in a loop--every day, and haven't hit any limits yet. It seemed to good to be true, after hearing others talk of $300/month bills. I guess it was.
$10/month was a too good of a deal.
The plan is to normalize spending hundreds/thousands on tokens per month for the productivity you gain.
See Jensen Huang’s comment that every $500k developer should spend at least $250k worth of tokens per year.
Glad this was announced because I didn't even realize that our (small) team has been paying $20/month for Github Copilot when none of us are using it, so used the opportunity to cancel CoPilot altogether. I think it was free when I first activated it (this was also before Claude, Codex, Gemini, and while not that great, it was why not), and didn't realize it had switched to paid and bundled with our Github bill, and now per usage.
Which one is it:
1. Current models in fact do not solve coding.
2. You can simply wait for a ~year for open-source to catch up and run it locally.
Re 1: Current models don't solve coding. They are useful tool for it though.
Re 2: Open weight models seem to be less than a year behind proprietary ones, so sure, if you're willing to spend tens or hundreds of thousands of dollars on a super computer that you probably don't fully utilize instead of renting time on someone else's super computer for a lot less.
... Once again the Business accounts get all sorts of goodwill [0] and users get the shaft.
[0] - Last weeks changes limited my personal Copilot Pro account but not my Work one
What "goodwill"? It's just more "AI credits" for what will be a shit product in June.
People need to wake up and stop being surprised by these billing increases. I see it on every update of every model. This was all subsidized by VC and company money. Now they need a return and the prices will keep going up. Be glad that you took advantage of that up until now, but can we stop the pearl clutching when we all know the amount of money being dumped into AI and the lackluster returns?
It's less surprise, but more confusing given the game theory as their competitors are not doing the same thing and the multiplier changes alone will likely churn current users.
Their competitors aren't doing the same thing because none of their competitors had the same broken 'pay per request' pricing model.
AI companies are increasing prices across the board.
"The bottom line" is the new "Its not just X, its Y"
Github Copilot has for the last week been an absolute shitshow. Clearly something has happened behind the scenes.
AI itself clearly weaker / dumber. Prices increased manyfold.
aaaand another github outage today. FFS
TLDR: It's a 6-9x price increase
yes...need to move to api usage, that a huge jump after getting everyone onboard
tldr: people were running multi-hour agentic coding sessions for the same flat fee as a one-liner autocomplete, github was eating the bill, and that party's over on june 1st
Still no chips coming from Microslop
Google won