Has either of these companies released models before this? It's hard to believe that they could release a supposed Mythos-level model just out-of-the-blue. Deepseek, Z.ai, Alibaba/Qwen have been at this for a lot longer and have been releasing models with steadily increasing capabilities for about 18 months now. I find it hard to believe that these new companies would just suddenly release a Mythos-level model without releasing anything prior.
asian is bad wording. this is a japanese startup backed by khosla ventures. japan is an ally of west.
the title makes it sound like a chinese company did this.
> These companies providing tokens, whether SOTA or not, that want to IPO are so fucked as time goes on.
>Can't sell their SOTA models, only slightly better than the open source models for the models they can sell, cost 20x to 50x for good models, a TAM that consists almost solely of developers, with no customer of theirs actually boasting increased profits as a result of AI...
> I fear their time to IPO may have passed.
What on earth could Anthropic and OpenAI Pivot to now?
The open models are half the equation. The other half is Apple's hardware, which is likely to see major memory bandwidth improvements over the next 2-3 generations and will be capable of running substantial models locally. By that point the open models will be beyond today's SOTA.
They may not get the valuation they want, but as it appears to be on a plateau may be better to offload now?
As per SpaceX, so many big names are involved the media will be controlled to hype it up and the investment banks will forecast 100x revenue in 2 years...
I agree with everything you said about their situation, but it's not like that is what will be evaluated in an IPO. There will be continued hype by the companies, lobbying to win support of a corrupt administration, and a narrative spin by clueless media about this AI revolution that will give investors fomo.
First impression: Third-party benchmarks or gtfo. Personally, I've never heard of either of these companies before. We're just supposed to take their word that they've matched the best models on the market?
Sakana describes their model as a "Orchestration Model." Does that mean that it's actually a bunch of different models glued together?
My impression is that the answer is yes, that it purports to dispense the glue on-the-fly in some kind of dynamic way rather than being some kind of new model-amalgam.
I think it is time that we had a UN-sponsored standards body dedicated to bench-marking the newest models from around the world, for everyone's benefit.
Given the national security implications, it's no surprise that Japan and China are rushing to build sovereign models post-ban.
But when these startups claim parity with "Mythos," could it be that they are just optimizing for very specific inference tasks?
I wonder if we are seeing the real battleground shift from raw training scale toward specialized inference.
Well if they are hyped like Mythos then we can add that to the list of “like Mythos”. Perhaps what’s missing is their CEO warning the world that their model is too unsafe to be released on the internet and someone must stop them before it’s too late.
I don't even look at benchmarks anymore. I just try different models as they're released on our large, proprietary, systems software codebases in real, shipping products or projects that will ship eventually. It's pretty clear which models help me do my job better or faster. I'm fortunate enough to have the token budget to use basically as much as I need, for now.
No need for benchmarks, evals, marketing, system cards or anything like that. I read the web for tips, practices and release announcements. My colleagues and I share our experiences with each other but beyond that, everything else is just noise.
I tried the Fugu models with some real world tales in C# and unity using mcp and open code. I exhausted the $20 plan 5 hour window in one prompt to review my theme system and plan some color changes. So I upgraded to the $100 to see the implementation and result. Well the result was worse than Opus, incredibly slow, and I ended up exhausting the new 5 hour window and have used 35% of the weekly now and it hardly created something opus was able to do at a fraction of the time and cost.
Do what you wish with this info, but it seems to be a complete waste of $$.
We provide a similar service for Godot instead of Unity, and 20$ plan being exhausted in one prompt on a top model like Opus sounds about right. That's the life when you pay API prices and can't afford 10x subsidies.
I tested Fable through Cursor; asked for ideas on how to make a data website I have less "Claude-like" (IYKYK what are the usual tells), and it spun out the most useless, Claude-like CSS styling ever, wasting $40 in 10 minutes.
The website was created through Opus, so you could also say the results were worse than Opus. (This is just to say that I had the same experience using the US models, so perhaps those Asian models are Mythos-like lol)
I experienced the same exact thing, however, I will say that I had misconfigured it on `pi` at first.
I was using its chat endpoint not the responses with tool calling and everything, and I haven't tried it again since, learned that recently and I am planning on giving it another shot.
This is useful info. For the couple of days that Fable was live - it was clearly a step above Opus 4.8 and I was able to get 8-10 prompts in using my $20 plan.
Because Fugu is not an independent model. They just use multiple existing SaaS models from OpenAI, Anthropic etc in background, gather response and generate results based on these response.
They claims that combining the results of multiple AI models and generate final result by using their in-house proprietary model improve the quality than using the single backend model.
It cause all sorts of doubts like: is their in-house model really exists? Is their in-house model really capable?
Personally, even if their claim is correct, such feature can be easily implemented in client side like Claude Code etc, using the equally capable model from background model to generate the final result.
I'm expecting a ban of "foreign" llms due to "safety concerns" before the year is over.
It will have nothing to do with the actual performance. But anthropic has set the bar for mythos-like systems, and whatever meets that loosely defined bar will be unsafe for the public.
Fugu Ultra [0] is not actually a model, it's a system (harness in the cloud?) that routes to several models, looks like it's a bit like OpenRouters Fusion [1].
"Rather than a single monolithic model, Fugu is a learned multi-agent orchestration system: a language model trained to route tasks across a swappable pool of underlying models and to recursively call instances of itself." - https://openrouter.ai/sakana/fugu-ultra
I doubt it will rival Mythos or the upcoming Sol, and if it's not open weights it doesn't really matter in the grand scheme of things. Still, I applaud the asian LLM efforts and hope they keep up the pressure on the americans.
unless they launched 10t param models, or figured out some amazing new way to compress as many params into say 100b, I doubt it's anywhere near "mythos level". and I have no idea how many params mythos has but that was just some hear say.
Competition is accelerating, but the next breakthrough isn't just better models it's better connectivity. AgentKey bridges AI agents with real-world tools, APIs, and data.
Just like many comments have been saying here, I also tested Fugu and some others and what I noticed is that they are quite expensive models, 20$ is not enough to complete a full workflow which in Opus it's possible, sure you might need to improve your prompt from the get go with Opus if you want the best results but so far that's my experience.
My next test will be Agentic systems and see how they perform
49 comments
[ 2.8 ms ] story [ 53.2 ms ] threadIt was bound to happen soon.
we're increasingly irrelevant
The idea here is: if you have a substantive point, make it thoughtfully; if not, please don't comment until you do.
https://news.ycombinator.com/newsguidelines.html
1. https://sakana.ai/company-info/?lang=en
2. https://news.ycombinator.com/item?id=48624782
https://techcrunch.com/2025/02/21/sakana-walks-back-claims-t...
> These companies providing tokens, whether SOTA or not, that want to IPO are so fucked as time goes on.
>Can't sell their SOTA models, only slightly better than the open source models for the models they can sell, cost 20x to 50x for good models, a TAM that consists almost solely of developers, with no customer of theirs actually boasting increased profits as a result of AI...
> I fear their time to IPO may have passed.
What on earth could Anthropic and OpenAI Pivot to now?
They may not get the valuation they want, but as it appears to be on a plateau may be better to offload now?
As per SpaceX, so many big names are involved the media will be controlled to hype it up and the investment banks will forecast 100x revenue in 2 years...
Sakana describes their model as a "Orchestration Model." Does that mean that it's actually a bunch of different models glued together?
See also contemporaneous reaction at:
https://news.ycombinator.com/item?id=48624782 (6 days ago, 244 points, 133 comments)
But encouraging for Japan to announce competition along with China.
No need for benchmarks, evals, marketing, system cards or anything like that. I read the web for tips, practices and release announcements. My colleagues and I share our experiences with each other but beyond that, everything else is just noise.
Do what you wish with this info, but it seems to be a complete waste of $$.
The website was created through Opus, so you could also say the results were worse than Opus. (This is just to say that I had the same experience using the US models, so perhaps those Asian models are Mythos-like lol)
Missed half the stuff the other half was outdated/ didn't verify.
I was using its chat endpoint not the responses with tool calling and everything, and I haven't tried it again since, learned that recently and I am planning on giving it another shot.
Ran into a package conflict issue with the popular coplay one
They claims that combining the results of multiple AI models and generate final result by using their in-house proprietary model improve the quality than using the single backend model.
It cause all sorts of doubts like: is their in-house model really exists? Is their in-house model really capable?
Personally, even if their claim is correct, such feature can be easily implemented in client side like Claude Code etc, using the equally capable model from background model to generate the final result.
I smell something fishy on their service.
It will have nothing to do with the actual performance. But anthropic has set the bar for mythos-like systems, and whatever meets that loosely defined bar will be unsafe for the public.
My next test will be Agentic systems and see how they perform