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)
Omitting those characters makes it good for generating passwords if they need to be typed in by hand.
Double-clicking a base58 string always selects the whole string and it doesn't wrap accidentally, thanks to missing / and +, so it's also convenient to copy and paste.
$20/month: Claude Code
$10/month: Minimax
$16/month: Xiaomi Mimo
$10/month: Opencode Go
Opus at low/medium effort generates plans. Then several coordinator/worker pairs are possible: DeepSeek v4 Pro + Minimax M3, Mimo v2.5 Pro + Mimo v2.5, Mimo + Minimax, Sonnet 4.6 + Haiku. I've been running hundreds of long multi-agent sessions, topped up extra credits here and theere, but haven't reached $200/month spend yet. Relying entirely on Claude/Codex feels like a waste of cash now.
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
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.
But it's a hefty upfront investment for people who just want to experiment. The good thing about $200/month subscriptions is that you can cancel them any time and cut your losses. Not so with a $4000 computer that loses half of its resale value as soon as you plug it in.
I think the current sweet spot for people who don't already own a high-end gaming PC is to rent a server with a beefy GPU from Hetzner et al. and run local models there.
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.
But, it is not all about cost: models like DeepSeek v4 flash (I use the US company Fireworks.ai and also buy tokens directly from DeepSeek) is very fast, very low latency while working.
Would you want to use a text editor that updates the screen very slowly? Kind of the same thing for using agentic systems as coding assistants: don’t want a ‘sluggish’ experience.
I have, mostly, long running autonomous tasks, so it doesn't matter how slow inference is. If I optimize for latency it means I'm turning into the limiting factor.
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
The beauty of your approach: when people are not paying for an expensive subscription, they can decide to use models less and not feel like they are leaving money on the table.
Redesigned and rearchitected our company's backend and mobile apps in about 2 months with 1 engineer and 1 PM trained to work with Claude. We used a custom-built multi-agent harness to crank through lots of parallel tasks, which necessitated the API usage.
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.
Indeed. The world models research many labs are now chasing was to some degree ignited by David Ha and Schmidhuber's 2018 paper.
More broadly, Sakana is pursing a refreshingly distinct research path, with their focus on evolutionary methods, biological intelligence (e.g. continuous thought machines) and open publication.
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.
I tried running this for some market research for my startup and it did a pretty nice job. It didn't necessarily find any obscure data, and it seemed to rely on older data than what I could find myself. On top of this, it had the same sycophantic tendencies as most LLMs these days (explaining why your idea is great and riffing on that), which I find to be unnecessary use of resources.
All put together, paying ~$60 to get a hit-or-miss report seems a bit excessive, but obviously as the models they use under the hood get better it becomes more and more worth it, assuming they also improve their grounding/search capabilities.
I'm a big fan of Sakana though, and have followed David Ha / @hardmaru since the world models papers (with the racing car game and the Doom clone), which were incredible at the time.
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 :-)
> 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?
Got myself the $20 subscription and tried it out. The 5-hour limit runs out surprisingly fast. Quality is okay but it feels slow, and even with my $20 Claude subscription on Fable, the credit usage ends up being lower. Fable usually catches issues in my Opus 4.8-generated code that I'd miss otherwise, but Fugu didn't. Makes me wonder if it's really at the Fable level. Hard to see the value here.
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.
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"...
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.
Imho there are two dimensions here: Firstly different LLMs and secondly the strategy in which you break down the problem in an agentic fashion (e.g. break up to separate agents with own persona and then judge evaluates across all agents). You can of course mix-up the dimensions as well and that's what I have been tinkering* with for a good few months with some success. This was all done using home-brew setup running on openrouter.
Personally I prefer understanding the dimensions and the interplay and controlling it though can see why openrouter and others are now offering this a solved solution.
Just be careful when you start outsourcing too much of your intelligence needs to a blackbox.
This is interesting. Would you share a few ways in which you're using this in your workflow? What about if you were to start a new project and test and built it out from scratch - how do you work this approach in without bogging everything down(including the simple things) down with overanalysis?
I only use this for high value problems/challenges. A lot related to life decisions including work, where to live, finances etc. It gives me a small army that can break down and slice, dice the problem in different ways then someone to reconcile it all and present it back to me. The variance in their opinions is the most interesting part of this project so f ar.
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.
I’ve been working on my own harness / “orchestration layer”, not with the goal of reaching frontier level performance, but rather boosting performance of smaller (locally hostable) models. Unfortunately, I don’t have VC money to burn on running hundreds of evals, but some preliminary results do indicate that it could work[0].
Nice idea but expensive. It looks like they don’t add very low cost models like DeepSeek v4 flash into their mix.
After a few months of spending money on the best frontier models, now I am spending time using DeepSeek v4 flash as my workhorse, and flipping to more capable (but still very inexpensive) open models on an as-needed basis. We all make our own tool selection decisions, but for me, I feel happier and enjoy working more following the very fast response and ultra low cost path.
Brilliant. What this actually is, is a swarm, albeit a very small one. I'm wondering if for research specifically, swarm size (on higher temp?) would outweigh model size.
At least, for the initial data gathering phase. You'd probably want a sequence of progressively larger models to filter it.
Have you guys tested it on anything other than research?
For anyone finding this, I used this during the beta. Beats GPT-5.5 xhigh on complex tasks. Since it’s expensive and difficult to subsidize, use it for the most challenging problems.
OAI/ANT can subsidize their own subscriptions, so it’s hard to compete there. But the results I got from fugu-ultra were impressive.
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.
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.
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!
See also: OpenRouter Fusion, similar idea, although it seems limited to internet research tasks? (Unclear, maybe someone who has used it can elaborate.)
What's nice is that OpenRouter included a pareto graph showing the cost as well as the performance. (But not time, unfortunately -- model fusion adds a large factor to round trip time.) Benchmarks are a lot less helpful without that.
OpenRouter: Surpassing frontier performance with fusion (blog post with benchmarks)
I did my own last weekend in a few lines of Python, though I haven't tested it much yet. (Looking for some very hard, very cheap benchmarks, if such a thing exists!)
i've seen many AI models, tried some. i'm genuinely interested in trying this kind of model/architecture. however i'm a little confused about the pricing.
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.
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.
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)
More fun facts:
Omitting those characters makes it good for generating passwords if they need to be typed in by hand.
Double-clicking a base58 string always selects the whole string and it doesn't wrap accidentally, thanks to missing / and +, so it's also convenient to copy and paste.
Unfortunately, no special characters means that a base58 string will often be rejected as a secure enough password.
My current setup:
Opus at low/medium effort generates plans. Then several coordinator/worker pairs are possible: DeepSeek v4 Pro + Minimax M3, Mimo v2.5 Pro + Mimo v2.5, Mimo + Minimax, Sonnet 4.6 + Haiku. I've been running hundreds of long multi-agent sessions, topped up extra credits here and theere, but haven't reached $200/month spend yet. Relying entirely on Claude/Codex feels like a waste of cash now.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
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.
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.
Sure, if you're going to keep using it long term.
But it's a hefty upfront investment for people who just want to experiment. The good thing about $200/month subscriptions is that you can cancel them any time and cut your losses. Not so with a $4000 computer that loses half of its resale value as soon as you plug it in.
I think the current sweet spot for people who don't already own a high-end gaming PC is to rent a server with a beefy GPU from Hetzner et al. and run local models there.
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.
The base model where I live (Central Europe) is $194. The Pro is $357. The Max is $779.
I just averaged it out.
But, it is not all about cost: models like DeepSeek v4 flash (I use the US company Fireworks.ai and also buy tokens directly from DeepSeek) is very fast, very low latency while working.
Would you want to use a text editor that updates the screen very slowly? Kind of the same thing for using agentic systems as coding assistants: don’t want a ‘sluggish’ experience.
I have, mostly, long running autonomous tasks, so it doesn't matter how slow inference is. If I optimize for latency it means I'm turning into the limiting factor.
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
The beauty of your approach: when people are not paying for an expensive subscription, they can decide to use models less and not feel like they are leaving money on the table.
Where do you acquire free hardware and free electricity such that you can host local models for $0?
Not everyone can run local models. It is also expensive will be outdated soon as the model evolves.
I wish I only paid $200/mo for Anthropic! Multiply that by 20x.
What are you getting out of it at $4000/month?
Redesigned and rearchitected our company's backend and mobile apps in about 2 months with 1 engineer and 1 PM trained to work with Claude. We used a custom-built multi-agent harness to crank through lots of parallel tasks, which necessitated the API usage.
i burned ~20k+/mo on codex.
Did you make those money back?
It is hard to stretch every single token to a win but …
The major two deals it was purposed to are still up on the air , if we win sure , 60x win
Does it work? I’m less interested in economics than fit with an MVP.
https://news.ycombinator.com/item?id=48625727
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.
Oh! I thought TrustedRouter was a joke/sarcasm. Very wrong placement of the comment.
It’s all open source and I say that it’s mine in all the sibling comments above
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.
Which software are you using to do that?
Edit: nevermind, but which plugin or so?
[dead]
Yo dawg, I heard you like agents, so we put agents in yo agents so you can burn tokens while you burn tokens.
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.
Indeed. The world models research many labs are now chasing was to some degree ignited by David Ha and Schmidhuber's 2018 paper.
More broadly, Sakana is pursing a refreshingly distinct research path, with their focus on evolutionary methods, biological intelligence (e.g. continuous thought machines) and open publication.
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.
Which model is that? What is it named?
Not OP but I believe it’s Fugu (7B). According to [0]: “Fugu itself is a trained coordinator LLM.”
[0] https://dev.classmethod.jp/en/articles/sakana-fugu-ga-first-...
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
FWIW, I've just published this interview with David Ha an hour ago.
We talk in detail about Fugu and why these kind of routing models are likely to win out over the big frontier models.
He makes some very good arguments for them.
https://www.disruptingjapan.com/the-future-of-ai-looks-very-...
I tried running this for some market research for my startup and it did a pretty nice job. It didn't necessarily find any obscure data, and it seemed to rely on older data than what I could find myself. On top of this, it had the same sycophantic tendencies as most LLMs these days (explaining why your idea is great and riffing on that), which I find to be unnecessary use of resources.
All put together, paying ~$60 to get a hit-or-miss report seems a bit excessive, but obviously as the models they use under the hood get better it becomes more and more worth it, assuming they also improve their grounding/search capabilities.
I'm a big fan of Sakana though, and have followed David Ha / @hardmaru since the world models papers (with the racing car game and the Doom clone), which were incredible at the time.
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 :-)
> 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?
Got myself the $20 subscription and tried it out. The 5-hour limit runs out surprisingly fast. Quality is okay but it feels slow, and even with my $20 Claude subscription on Fable, the credit usage ends up being lower. Fable usually catches issues in my Opus 4.8-generated code that I'd miss otherwise, but Fugu didn't. Makes me wonder if it's really at the Fable level. Hard to see the value here.
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.
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"...
See also: Agents built from alloys (July 2025)
https://news.ycombinator.com/item?id=44630724
They randomly alternated between frontier LLMs and got a massive boost to performance on cybersecurity tasks.
I love when they put a black box in front of the other black boxes so I can get a questionably better black box for slower service and more money!
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
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
Fugu, eh? So there’s a nonzero chance this thing might kill me?
Imho there are two dimensions here: Firstly different LLMs and secondly the strategy in which you break down the problem in an agentic fashion (e.g. break up to separate agents with own persona and then judge evaluates across all agents). You can of course mix-up the dimensions as well and that's what I have been tinkering* with for a good few months with some success. This was all done using home-brew setup running on openrouter.
Personally I prefer understanding the dimensions and the interplay and controlling it though can see why openrouter and others are now offering this a solved solution.
Just be careful when you start outsourcing too much of your intelligence needs to a blackbox.
* https://github.com/monkeydust/rightmind
This is interesting. Would you share a few ways in which you're using this in your workflow? What about if you were to start a new project and test and built it out from scratch - how do you work this approach in without bogging everything down(including the simple things) down with overanalysis?
I only use this for high value problems/challenges. A lot related to life decisions including work, where to live, finances etc. It gives me a small army that can break down and slice, dice the problem in different ways then someone to reconcile it all and present it back to me. The variance in their opinions is the most interesting part of this project so f ar.
Thanks. I like the idea of doing this, myself, and glad you see a value from this.
Do you not worry about giving away your most intimate data to for-profit companies who have not signed to protect your data in a dignified fashion?
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
I’ve been working on my own harness / “orchestration layer”, not with the goal of reaching frontier level performance, but rather boosting performance of smaller (locally hostable) models. Unfortunately, I don’t have VC money to burn on running hundreds of evals, but some preliminary results do indicate that it could work[0].
https://deepclause.substack.com/p/how-to-make-small-models-p...
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.
Nice idea but expensive. It looks like they don’t add very low cost models like DeepSeek v4 flash into their mix.
After a few months of spending money on the best frontier models, now I am spending time using DeepSeek v4 flash as my workhorse, and flipping to more capable (but still very inexpensive) open models on an as-needed basis. We all make our own tool selection decisions, but for me, I feel happier and enjoy working more following the very fast response and ultra low cost path.
We found that an all open source fusion was 1/3 the price and better than Fable
https://trustedrouter.com/blog/open-fusion-beats-fable-5
Brilliant. What this actually is, is a swarm, albeit a very small one. I'm wondering if for research specifically, swarm size (on higher temp?) would outweigh model size.
At least, for the initial data gathering phase. You'd probably want a sequence of progressively larger models to filter it.
Have you guys tested it on anything other than research?
Working on bio coding and cybersecurity benchmarks now
For anyone finding this, I used this during the beta. Beats GPT-5.5 xhigh on complex tasks. Since it’s expensive and difficult to subsidize, use it for the most challenging problems.
OAI/ANT can subsidize their own subscriptions, so it’s hard to compete there. But the results I got from fugu-ultra were impressive.
Sakana is certainly a choice of a name. In portuguese, goes from anything between "scoundrel" to "sleazebag".
How do you configure it to run with pi or claude code? I'm curious to try it (via subscription ideally)
EDIT: Found something here https://dev.classmethod.jp/en/articles/sakana-fugu-ga-first-...
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.
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.
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?
See also: OpenRouter Fusion, similar idea, although it seems limited to internet research tasks? (Unclear, maybe someone who has used it can elaborate.)
What's nice is that OpenRouter included a pareto graph showing the cost as well as the performance. (But not time, unfortunately -- model fusion adds a large factor to round trip time.) Benchmarks are a lot less helpful without that.
OpenRouter: Surpassing frontier performance with fusion (blog post with benchmarks)
https://news.ycombinator.com/item?id=48525392
OpenRouter Fusion API
https://news.ycombinator.com/item?id=48537641
See also: Sibling comment with an open source implementation
https://news.ycombinator.com/item?id=48624782#48629598
I did my own last weekend in a few lines of Python, though I haven't tested it much yet. (Looking for some very hard, very cheap benchmarks, if such a thing exists!)
i've seen many AI models, tried some. i'm genuinely interested in trying this kind of model/architecture. however i'm a little confused about the pricing.
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.
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?
So plain old ensemble technique in classical ML.
Is this just OpenRouter Fusion but from Japan?
$40/month on Kiro, I hardly ever hit my credit limit
Just letting you guys know that the model is not a moat.
Is this the beginning of the Hyperion TechnoCore?
https://xkcd.com/927/
And it's gonna be like that in this space for the next decade-plus.
And yet, as per usual...
The price to pay for claiming your laws are applicable worldwide.
[dead]
[dead]
[flagged]
[flagged]
[dead]
[dead]
I’ve also developed and open-sourced Mythos level model using fusion/synthesis on TrustedRouter
https://trustedrouter.com/blog/fusion-evals-open-source
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.