You pay $200/month to Anthropic, $200/month to OpenAI, $200/month to Cursor, $200/month to $200/month to Google, and seeing that it didn't come to a nice round $1024/month, you pay $200/month to Sakana to coordinate it all, because why not.
While you're at it, feel free to send me $200 as well, I'll generate a crypto address ending with "AI".
TIL: I just found out that base58 disallows I (capital i), l (lowercase L), O (capital o) and 0 (zero), so I could only generate GrxoJt4eNXE2QaQ55iPSa7hhiYdzCo8ZeAuokmh2Cai.
(don't send anything, sharing only because of the base58 fun fact I didn't know)
at this point I might just try Neuralwatt and see how much request I can get with GLM5.2. I've read a lot of reviews that its very cheap to run using Neuralwatt cloud
I used to have a $20/mo ChatGPT subscription and now I spend $12 per year using Kimi models on OpenRouter, and that's with zero-data-retention-only providers (some models sometimes have free providers with scary tracking). Maybe I just don't use that many tokens, I don't fill the context with more than what's needed for a specific request, but it goes to show how these subscriptions can be an absolute ripoff. The thought of spending 200x that is insane to me
Of course, premium headphones existed before. I have a WH-1000XM4 sitting right next to me.
But your aunt Josie didn't have one. Now Apple is selling 80 million units / year and the ~$300 price tag has become normal. Before that, most people had headphones that were 10 times cheaper.
$300 isn’t what AirPods cost though. You can get a pair of AirPods 4 for $129 on Apple.com, and I presume that is still the most popular model. If you’re paying ~$300, you are buying premium headphones.
Not while the hardware required to run a local model at an acceptable speed costs way more than $200.
Guess what, the big players are hoarding all the RAM and GPUs so that other people can't afford decent hardware. It's working out beautifully for them!
> Not while the hardware required to run a local model at an acceptable speed costs way more than $200
It's $200/month. You have to take into account energy costs and all the rest of a system, but if you break even within 1-2 years ($2400-$4800) it'd be a pretty good deal. And $4000 buys you a pretty decent system.
Happy user here, pairing it with Composer 2.5, with Fugu Ultra as advisor and Fugur as planner. For scope/architecture it’s on par with useful Fable-style orchestration than one chat thread.
I've been shipping production on archive.tw with Fugu Ultra in /advisor on oh-my-pi.
Advisor doesn’t slow the loop if the driver stays fast. Worth it if your harness can split advisor from worker.
David Ha, CEO and co-founder, was one of the youngest managing director at Goldman Sachs before doing ML at Google. His ML publications were considered top-notch almost a decade ago. I had high hopes for him when he raised money and founded Sakana.
I do agree with some comments here that perhaps this particular product is not well thought out. I also agree with the criticism that David calls Sakana a frontier AI lab while making money just selling AI B2B applications to Japanese businesses. I also agree with the assessment that Sakana has abrasive and antagonistic, sometimes openly hostile, recruiting tactics. I also agree that his then-impressive publications may have lost their luster in the age of LLMs.
However, the man is clearly driven; and he and his team may have more to offer in future. I admire the man for not taking the conventional AI-research career path.
Looking at the technical report I'm a bit confused. The improvement from using their orchestrator models seems minimal (in some cases lower than just the model which I'm assuming is in the orchestrator's pool?). Maybe it's sort of acting as an additional reasoning step upfront? Sort of like how if you asked Claude to create a plan for how best to prompt itself, you would probably end up with a better result than just the base prompt.
Also, from the technical report, looks like they're training on the output of Claude Code, etc. I'm guessing this doesn't violate TOS because they're technically not a directly competing model. This brings me to what I see as the main risk with this service, which is that it seems like an easy thing for a frontier lab to make obsolete, either by models beginning to converge in terms of strengths or by improving their own harnesses to include more of this meta-reasoning.
As a developer outside the US I think it's vital to have alternatives to OpenAI and Anthropic, but sadly this is not it. For $200/month you get < 3 hours of use per week, the API is extremely slow, and the output quality in my tests is nowhere near Fable. It's nowhere remotely near usable as a day-to-day workhorse. Very disappointing.
u seem to be the only one who used it here - how did it compare to opus and gpt5.5? in theory it should be at least on par if not better at times right.
I only had time to use it for a couple of deep reviews of large Rust projects, and a few agentic coding tasks (implement plan X, refactor Y in fashion Z) before my quota ran out. My impression is that the reviews were quite strong - maybe Opus 4.8+ or around GPT 5.5 (for my particular use case) - but very slow. For implementation I found it weaker, it made a few mistakes that I haven't seen frontier models make in a long time.
ngl, I thought sakana.ai was doing cooler stuff than this. that said, the release of a product like this makes sense because it follows your natural intuition when using these models. The best way to use LLMs is to have at least two in your pocket, because the models do a good job at covering each others assets and filling in obvious model-specific blindspots.
it's interesting that they're offering in the form of fixed cost subscription plans too. My impression was that the first party providers can do this because they api inference margins to the tune of 80ish percent. Anyone else orchestrating on top of these models have to pass through these costs or eat it themselves.
> Frontier-level performance without single-vendor dependency. [...] Plug collective intelligence directly into your workflows today with a single API.
Does multiple vendors run this "single API" or how is this not replacing a single-vendor dependency for another single-vendor dependency?
Beta user: they piloted OpenRouter fusion before it was seen as the viable step. Everyone's understood for months now that having different models check each other is the best path forward.
This gets you that in a nice neat package, without the underlying tinkering mechanics.
If (big iff) the usage mechanics work out, then this is actually a really good anti-big-model strategy.
They'll be incentivized for your success, not token-maximizing for their investors.
if you've used codex or claude, how do the usage limits on fugu feel compared to the pro plans on either? honestly wouldn't mind subscribing to this if it's as generous as what codex is giving me monthly, which seems unrealistic.
Hard to say - since I used it in Beta with free credits, where the usage felt more 'Opus' than 'ChatGPT' but more efficient token wise. Switching models every time is annoying.
But their paid plans I'm not sure yet - planning to subscribe and can let you know.
Almost no chance it will be as generous as OpenAI though. They just don't have the money :-)
Is there any official source that could confirms if Fable (or Mythos) is parallelized test-time compute (like GPT 5.5 Pro) or sparse Mixture-of-Experts (MoE) transformer combined with a multi-agent, inference-time compute scaling architecture (Gemini 3.1 Deep Think)?
Not necessarily. There were some tests last year-ish from hf that showed that simply alternating (randomly) between claude and gpt (whatever their versions were at the time) on a task produced better results than either of them individually. So during a task, the first call was sent to one, then the other and so on.
There's also the concept of "smart routing" requests based on some heuristics / embeddings. You'd get "simple" tasks handled by smaller (cheaper) models and use a bigger model to curate / sort / merge the results.
There's a lot of things to try here. I wouldn't personally pay for this service, but I don't think it's "a joke"...
OpenRouter Fusion is basically ask N models + synthesizer step.
This is ask a special orchestrator they built, which is in front of a bunch of models, which model would suit the request best.
Regular Fugu seems to be just "pick the best model and route the request there"
Fugu Ultra can generate like a little mini workflow/plan instead to achieve a result
1. Ask GPT to derive the math.
2. Ask Opus to check for implementation/security issues.
3. Ask Gemini to synthesize or resolve disagreement.
4. Return final answer.
I could be wrong but seems to be that at a glance, so I think it's more dynamic than OpenRouter Fusion.
Yeah, I was trying to parse their "defense policy"
https://sakana.ai/company-info/defense-policy.html?lang=en
But it seems like lot of words to say we have no policy and we'll just go along with the powers that be. Like they rely on deferring to the Pacifist constitution, which the current administration if moving mountains to try and change. And when it it you can bet they will not want to give up their defense contracts.
I was just stating facts about Sakana, and that was enough to trigger you? For the same reason, I don’t use GPT either. At least for now, DeepSeek has no ties to the defense sector. And don’t talk as if the CCP were the devil. The U.S. president is the world’s biggest arms dealer, after all.
Their research around building a domain specific model is pretty cool, it's kind of like Karpathy's autoresearch but pointed at deciding the optimal model to use at each step of the inference.
If cost becomes an even bigger problem being able to choose "best performance possible" or "strong but cost effective" will be useful.
This would have been much more interesting and impactful if it had relied on open source models rather than commercial models that are only availble via an API.
The reasoning chains could have been used, and the resulting combined model could easily and effectively have been distilled.
Looks like Fusion calls a bunch of models and then uses an LLM to synthesize the results, and pass to another model for final output.
Fugu looks like it's doing something different? Using an LLM earlier on in the flow as an orchestrator to decide which other LLMs to call. More coordinator than simply synthesizing results, and more "agentic".
It's interesting because it's all exposed behind a single OpenAI compatible endpoint (Responses API?) and so then presumably someone could use this for one of their single agents. Now you have agent-of-agents, nested in some sense. The token usage increases accordingly!
Basically, if you combine a bunch of near-frontier models (like GPT 5.5, etc) you can get performance that sometimes surpasses top line models like Claude's Fable.
Sakana seems to have a separate approach using a domain specific model to perform the model routing step.
This is a charitable read, but I think that being able to pick from a panoply of models will actually yield much better results in the long run.
The same model that has been post-trained to operate for hours as a Linux admin will be incapable of writing a heartfelt email, but with something like Fugu, you'd get both the Linux admin for driving the browser harness and the smaller writing specialist model for drafting the email itself.
AI noob question, is this like Amp? I just use Amp, I ask it to do neat stuff and it does it. I desperately need to invest in my AI skills but every day I open two new tabs and add it to "AI stuff" folder, and then go back to drowning in work to do.
You pay $200/month to Anthropic, $200/month to OpenAI, $200/month to Cursor, $200/month to $200/month to Google, and seeing that it didn't come to a nice round $1024/month, you pay $200/month to Sakana to coordinate it all, because why not.
While you're at it, feel free to send me $200 as well, I'll generate a crypto address ending with "AI".
TIL: I just found out that base58 disallows I (capital i), l (lowercase L), O (capital o) and 0 (zero), so I could only generate GrxoJt4eNXE2QaQ55iPSa7hhiYdzCo8ZeAuokmh2Cai.
(don't send anything, sharing only because of the base58 fun fact I didn't know)
Yo dawg, I heard you like agents, so we put agents in yo agents so you can burn tokens while you burn tokens.
Does it work? I’m less interested in economics than fit with an MVP.
https://news.ycombinator.com/item?id=48625727
at this point I might just try Neuralwatt and see how much request I can get with GLM5.2. I've read a lot of reviews that its very cheap to run using Neuralwatt cloud
Or use openrouter and switch to model you want to use..(i think so)
Or TrustedRouter if you want privacy and open source
You ought to realize that shilling your product in the comments doesn't exactly come across as trustworthy.
Disclosing affiliation hasn’t been a legal thing for a while. It’s reputational. Knowing that firm spams is a black mark.
Pay $0 to run a local model or even a cheap DeepSeek V4 model via their API which is close to free per million tokens.
These prices are just going to get raced to $0.
I used to have a $20/mo ChatGPT subscription and now I spend $12 per year using Kimi models on OpenRouter, and that's with zero-data-retention-only providers (some models sometimes have free providers with scary tracking). Maybe I just don't use that many tokens, I don't fill the context with more than what's needed for a specific request, but it goes to show how these subscriptions can be an absolute ripoff. The thought of spending 200x that is insane to me
Maybe. But for now it's fascinating how $200/month has kind of become a normal tier.
It's similar to how AirPods normalised all of us having $300+ headphones. All of us would have scoffed at the idea a decade ago.
Many people here spent a lot more than $300 on headphones long before AirPods appeared.
Those were hobbyists, audiophiles, professionals, artists (recording, performing, etc.).
They are talking about a much larger group of people.
I think OP meant noise-cancelling headphones, which were fairly ubiquitous in tech circles in open offices; before Apple launched AirPods.
Airpods Inc. would be very high up SP500 as a standalone business.
I had a really nice Sennheiser before that, too. But now you hop on the subway and everybody sports one.
The Sony WH-1000XM series and the Bose QC35 were the standard quality headphones years before AirPods were a thing, and both retailed at $300+.
Of course, premium headphones existed before. I have a WH-1000XM4 sitting right next to me.
But your aunt Josie didn't have one. Now Apple is selling 80 million units / year and the ~$300 price tag has become normal. Before that, most people had headphones that were 10 times cheaper.
$300 isn’t what AirPods cost though. You can get a pair of AirPods 4 for $129 on Apple.com, and I presume that is still the most popular model. If you’re paying ~$300, you are buying premium headphones.
Not while the hardware required to run a local model at an acceptable speed costs way more than $200.
Guess what, the big players are hoarding all the RAM and GPUs so that other people can't afford decent hardware. It's working out beautifully for them!
> Not while the hardware required to run a local model at an acceptable speed costs way more than $200
It's $200/month. You have to take into account energy costs and all the rest of a system, but if you break even within 1-2 years ($2400-$4800) it'd be a pretty good deal. And $4000 buys you a pretty decent system.
Happy user here, pairing it with Composer 2.5, with Fugu Ultra as advisor and Fugur as planner. For scope/architecture it’s on par with useful Fable-style orchestration than one chat thread.
I've been shipping production on archive.tw with Fugu Ultra in /advisor on oh-my-pi.
Advisor doesn’t slow the loop if the driver stays fast. Worth it if your harness can split advisor from worker.
There are so many derisive comments here.
David Ha, CEO and co-founder, was one of the youngest managing director at Goldman Sachs before doing ML at Google. His ML publications were considered top-notch almost a decade ago. I had high hopes for him when he raised money and founded Sakana.
I do agree with some comments here that perhaps this particular product is not well thought out. I also agree with the criticism that David calls Sakana a frontier AI lab while making money just selling AI B2B applications to Japanese businesses. I also agree with the assessment that Sakana has abrasive and antagonistic, sometimes openly hostile, recruiting tactics. I also agree that his then-impressive publications may have lost their luster in the age of LLMs.
However, the man is clearly driven; and he and his team may have more to offer in future. I admire the man for not taking the conventional AI-research career path.
so he's the quintessential brilliant jerk, ok
Kind of shocking - a model comes out that beats mythos and offers a reasonable price and it ... gets downvoted?
Probably taking hate from both sides - OpenAI / Claude fans who are undercutting its moat. Chinese open-model fans that want it to be cheaper.
But it's a genuine accomplishment to hit those benchmarks and offer a reasonable plan?
Bizarre reaction TBH.
Looking at the technical report I'm a bit confused. The improvement from using their orchestrator models seems minimal (in some cases lower than just the model which I'm assuming is in the orchestrator's pool?). Maybe it's sort of acting as an additional reasoning step upfront? Sort of like how if you asked Claude to create a plan for how best to prompt itself, you would probably end up with a better result than just the base prompt.
Also, from the technical report, looks like they're training on the output of Claude Code, etc. I'm guessing this doesn't violate TOS because they're technically not a directly competing model. This brings me to what I see as the main risk with this service, which is that it seems like an easy thing for a frontier lab to make obsolete, either by models beginning to converge in terms of strengths or by improving their own harnesses to include more of this meta-reasoning.
As a developer outside the US I think it's vital to have alternatives to OpenAI and Anthropic, but sadly this is not it. For $200/month you get < 3 hours of use per week, the API is extremely slow, and the output quality in my tests is nowhere near Fable. It's nowhere remotely near usable as a day-to-day workhorse. Very disappointing.
https://x.com/cortesi/status/2068898694238486658
u seem to be the only one who used it here - how did it compare to opus and gpt5.5? in theory it should be at least on par if not better at times right.
I only had time to use it for a couple of deep reviews of large Rust projects, and a few agentic coding tasks (implement plan X, refactor Y in fashion Z) before my quota ran out. My impression is that the reviews were quite strong - maybe Opus 4.8+ or around GPT 5.5 (for my particular use case) - but very slow. For implementation I found it weaker, it made a few mistakes that I haven't seen frontier models make in a long time.
I'm glad eager people like you test for lazy people like me
ngl, I thought sakana.ai was doing cooler stuff than this. that said, the release of a product like this makes sense because it follows your natural intuition when using these models. The best way to use LLMs is to have at least two in your pocket, because the models do a good job at covering each others assets and filling in obvious model-specific blindspots.
it's interesting that they're offering in the form of fixed cost subscription plans too. My impression was that the first party providers can do this because they api inference margins to the tune of 80ish percent. Anyone else orchestrating on top of these models have to pass through these costs or eat it themselves.
> Frontier-level performance without single-vendor dependency. [...] Plug collective intelligence directly into your workflows today with a single API.
Does multiple vendors run this "single API" or how is this not replacing a single-vendor dependency for another single-vendor dependency?
Beta user: they piloted OpenRouter fusion before it was seen as the viable step. Everyone's understood for months now that having different models check each other is the best path forward.
This gets you that in a nice neat package, without the underlying tinkering mechanics.
If (big iff) the usage mechanics work out, then this is actually a really good anti-big-model strategy.
They'll be incentivized for your success, not token-maximizing for their investors.
The team is super smart too. What's not to like?
Wishing them the best on launch.
if you've used codex or claude, how do the usage limits on fugu feel compared to the pro plans on either? honestly wouldn't mind subscribing to this if it's as generous as what codex is giving me monthly, which seems unrealistic.
Hard to say - since I used it in Beta with free credits, where the usage felt more 'Opus' than 'ChatGPT' but more efficient token wise. Switching models every time is annoying.
But their paid plans I'm not sure yet - planning to subscribe and can let you know.
Almost no chance it will be as generous as OpenAI though. They just don't have the money :-)
Reminds me of <https://github.com/irthomasthomas/llm-consortium>
Fugu Ultra <https://console.sakana.ai/models#fugu-ultra> sounds similar to GPT-5.5 Pro or Gemini 3.1 Deep Think .
Is there any official source that could confirms if Fable (or Mythos) is parallelized test-time compute (like GPT 5.5 Pro) or sparse Mixture-of-Experts (MoE) transformer combined with a multi-agent, inference-time compute scaling architecture (Gemini 3.1 Deep Think)?
This is a joke, right?
Not necessarily. There were some tests last year-ish from hf that showed that simply alternating (randomly) between claude and gpt (whatever their versions were at the time) on a task produced better results than either of them individually. So during a task, the first call was sent to one, then the other and so on.
There's also the concept of "smart routing" requests based on some heuristics / embeddings. You'd get "simple" tasks handled by smaller (cheaper) models and use a bigger model to curate / sort / merge the results.
There's a lot of things to try here. I wouldn't personally pay for this service, but I don't think it's "a joke"...
So basically... openrouter?
OpenRouter Fusion is basically ask N models + synthesizer step.
This is ask a special orchestrator they built, which is in front of a bunch of models, which model would suit the request best.
Regular Fugu seems to be just "pick the best model and route the request there"
Fugu Ultra can generate like a little mini workflow/plan instead to achieve a result
1. Ask GPT to derive the math. 2. Ask Opus to check for implementation/security issues. 3. Ask Gemini to synthesize or resolve disagreement. 4. Return final answer.
I could be wrong but seems to be that at a glance, so I think it's more dynamic than OpenRouter Fusion.
Then there is this as well.
https://www.databricks.com/blog/introducing-omnigent-meta-ha...
links to two papers with at least enough apparent quality and novelty to get into ICLR 2026
> So basically... openrouter
:skull:
i now really wonder how many people of the public understood my thesis defense lol
Will Le Chat try to eat Sakana? There is Le Chaton Fat and then there is Sakana Fugu too..
> Le Chaton Fat
For others looking around: LCF is a meme model, it's not real. It's a joke.
I probably will never pay to Sakana, as they are involved in military contracts.
https://japannews.yomiuri.co.jp/politics/defense-security/20...
Yeah, I was trying to parse their "defense policy" https://sakana.ai/company-info/defense-policy.html?lang=en But it seems like lot of words to say we have no policy and we'll just go along with the powers that be. Like they rely on deferring to the Pacifist constitution, which the current administration if moving mountains to try and change. And when it it you can bet they will not want to give up their defense contracts.
Anthropic was very much willing & involved in U.S. military contracts before the falling out with the DoD. OpenAI is actively involved.
https://openai.com/index/our-agreement-with-the-department-o...
I imagine if it was Deepseek partnering with the CCP it would be different?
I was just stating facts about Sakana, and that was enough to trigger you? For the same reason, I don’t use GPT either. At least for now, DeepSeek has no ties to the defense sector. And don’t talk as if the CCP were the devil. The U.S. president is the world’s biggest arms dealer, after all.
Of course DeepSeek is used by the PLA, why would you think otherwise? https://www.scmp.com/news/china/military/article/3303512/chi... There can be multiple devils, after all.
> At least for now, DeepSeek has no ties to the defense sector.
Like every company based in China they are under the control of the Chinese state, which is an armed entity known to use violence.
Their research around building a domain specific model is pretty cool, it's kind of like Karpathy's autoresearch but pointed at deciding the optimal model to use at each step of the inference.
If cost becomes an even bigger problem being able to choose "best performance possible" or "strong but cost effective" will be useful.
https://arxiv.org/pdf/2512.04695
This would have been much more interesting and impactful if it had relied on open source models rather than commercial models that are only availble via an API.
The reasoning chains could have been used, and the resulting combined model could easily and effectively have been distilled.
Seems kinda underwhelming considering they raised like $400M.
400m is the new 400k! Just look at the other company evaluations and how much they raised vs what they delivered
it's just one of their products right
Very interesting. I wonder if its kinda functions similarly to how OpenRouter's fusion API does. Hopefully isn't too long to respond.
Yea similar, possibly even more steps / slower. I put together an all open source fusion at 1/3 of price of Fable: https://trustedrouter.com/blog/open-fusion-beats-fable-5
We open sourced it all
and will be releasing a similar orchestrator next week on TrustedRouter
From a brief reading of what Fusion does: https://openrouter.ai/docs/guides/features/plugins/fusion
Looks like Fusion calls a bunch of models and then uses an LLM to synthesize the results, and pass to another model for final output.
Fugu looks like it's doing something different? Using an LLM earlier on in the flow as an orchestrator to decide which other LLMs to call. More coordinator than simply synthesizing results, and more "agentic".
It's interesting because it's all exposed behind a single OpenAI compatible endpoint (Responses API?) and so then presumably someone could use this for one of their single agents. Now you have agent-of-agents, nested in some sense. The token usage increases accordingly!
Isn't this what perplexity is?
Is Perplexity still a daily driver for a lot of folks?
Can someone explain this in layman terms? I don't understand any of it
It's similar to this: https://openrouter.ai/blog/announcements/fusion-beats-fronti...
Basically, if you combine a bunch of near-frontier models (like GPT 5.5, etc) you can get performance that sometimes surpasses top line models like Claude's Fable.
Sakana seems to have a separate approach using a domain specific model to perform the model routing step.
So plain old ensemble technique in classical ML.
But it's priced the same as frontier models. Why do I not directly pay for frontier models?
This is a charitable read, but I think that being able to pick from a panoply of models will actually yield much better results in the long run.
The same model that has been post-trained to operate for hours as a Linux admin will be incapable of writing a heartfelt email, but with something like Fugu, you'd get both the Linux admin for driving the browser harness and the smaller writing specialist model for drafting the email itself.
u get a pool of them + sakana?
I’ve also developed and open-sourced Mythos level model using fusion/synthesis on TrustedRouter
https://trustedrouter.com/blog/fusion-evals-open-source
https://xkcd.com/927/
Just letting you guys know that the model is not a moat.
AI noob question, is this like Amp? I just use Amp, I ask it to do neat stuff and it does it. I desperately need to invest in my AI skills but every day I open two new tabs and add it to "AI stuff" folder, and then go back to drowning in work to do.
And yet, as per usual...
The price to pay for claiming your laws are applicable worldwide.