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.
Is this true even for more targeted prompts that aren't about looking over an entire codebase or w/e? I just stick to my sub pricing and find good success on targeted requests, but I wonder if I would run up against things even then if not for 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)
If you actually give it an example of the style it can copy it well. even just screenshots of other websites or UIs. It just sucks at producing it itself.
Well, in part because the phenomenon has been discussed on Web forums that (a) have at this point made their way back into training data and (b) are accessible in Web searches that the model can invoke. And in part because the model can "know" what its initial instinct is and "decide" to go against it.
I tested Fable for a whole day. And my experience was quite the opposite. I was blown away. Admittedly, I did not try it through a middleman like Cursor. I used Claude Code CLI.
Me too. It’s output was fabulous. And it acted like a senior engineer - actually coding up hypotheses, testing them, finding problems and presenting good, usable recommendations backed by solid evidence and wisdom. It can probably do most of my job, which gave me a bit of an existential crisis.
I’ve paused my Claude subscription until they bring it back. Opus makes mistakes constantly, on every level of abstraction. Every time I look closely at its work I find problems.
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.
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
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
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.
This is the way. Not that big of budget here. But if there’s something promising, I just try that for a month or so. But even then… at this moment I’m using z.ai models and those do the job. No need for anything else. So I’m staying until there is something new, same affordability, but a lot better. (Using a coding plan)
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.
How do we define "mythos level" exactly outside of marketing buzz? I don't even think the majority of us can access Mythos yet even to make a comparison.
But did they deliver anything yet?
There are many startups with notable founders that are not being able to get models launched. And even if they are able to launch, so they get PMF.
In case of Sakana, they clearly focus on the Japanese market an have buildup a good pipeline on sovereign AI. But similar to Aleph Alpha or to a certain extend Mistral, I don't see how they can keep up.
I think every country/major country at the government level will back any home grown talent in view of USA restrictions.
The models don't need to be as good, just good enough for the task. I am using a Claude Q3 2024 model mostly (Haiku 4.5) at present, and it delivers what I need.
I live in Japan and yet can't seem to pay for their API in JPY... I bet their enterprise customers don't have that problem but it was pretty annoying given "AI in Japan" appears to be there only selling point.
The only way they can compete in Japan is the enterprise-game. Their partnership with Daiwa and MUFG is probably exactly what they should be doing. I doubt that they get that far though. Like Mistral who has partnerships with BNP Paribas and Airbus, they deploy in an on-premise or private cloud and in those settings, their models make good PoC's but cant compete where it gets interesting -- Workspace Agents. If you look at them from a Chatbot over RAG perspective, maybe they can do it. But that's tech from 2023.
Mythos is extreme hype. We are at a combo of authoritarian politicians peddling fear for power and tech bros trying to extract maximum investment returns.
We’re going to have many LLM/toolset combos that do what mythos does.
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.
It is materially better, but I didn’t feel a huge loss when it was yanked. My use case is large complex legacy modernization projects and its ability is definitely better than opus 4.8 at the job. But it’s more like an optimization, I could have a single or 2 pass in fable vs 8-10 with opus to arrive at the same solution.
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?
And even if they didn't, they have a track record. Even if we did have benchmarks in this case I would still wait until people got there hands on it and formed a more holistic opinion.
No, stop right there. Anything published by Anthropic implicitly is not third party. For it to be third party, the third party has to be the one publishing it.
Is it actually that hard to make good models or is it just about the amount of resources you have to do training? (This is an actual question, I really don't know.) I'm sure it's not trivial but does it really take world class secret knowledge to build off of the known existing techniques? I feel like there's tons of low hanging fruit still to explore, and time and resources are the limiting factor.
I suspect that Grok has been ironically lobotomized by pressures to correct its political views.
Similarly, I could imagine the Gemini folks working in a significantly more complex corporate climate, with different parts of Google pushing for different capability focuses. They are only lagging behind less than a year, so it isn't too large of a gap yet.
That said, the fact that Anthropic is currently the top dog suggests that talent and execution is incredibly important. A year ago none of my normie friends new them, and when i suggested using Claude looked at me like when I recommend Linux.
That shouldn’t affect Grok’ coding ability. How often are people discussing politics with Claude code? Writing decent code is just hard and it’s not just Grok.
If training a good model requires talent then that’s the answer to the question this thread is trying to answer: is training a good model actually that hard?
You’d be quite surprised, I think. Fine tuning a model on one axis can have drastic impacts on another that as a human we would expect to be completely unrelated.
Like Zuckerberg, top talent may not work with a polarising character if they disagree with his behaviour. Space focused talent don't have many choices aside from SpaceX but ai companies are a plenty and a top AI person can pick and choose.
If they have a top team and the money then appears to be a matter of a year or two? And one startup mentioned is Japanese not Chinese so they won't be banned from buying US tech.
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.
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.
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.
> 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?
Let’s face it - without the humans these machines ain’t shit - aka we have mechanically figured out ways to make machines better than us at certain things (on demand memory) but this idea they are intelligent is horse shit.
Btw the bar is low too! Most human created decks are garbage. And yet LLM’s don’t even beat those.
Machines have nothing to do with what speaks to the human - at best they are media to transmit what a human wants to express.
Having all the data in the world doesn’t solve that. Although if enough of society gets exposed to slop and there is an erosion of taste and quality… in fact I’m gonna stop there. It’s a dark path.
I used to agree with you but now do not. I now think the floor for this market is probably no worse than the annual revenue of cell phone plans in the US market. So say, $250 billion.
Now, that probably doesn't justify the valuations and hype being thrown around, but I think it gets at a real revenue number.
I also don't know how that number fits into the funding rounds already raised and VC dreams of IPOs for these two.
This isn't coming from deep analysis on a verifiable source, but I started asking people in my social circle (includes white-collar and blue-collar folks) about their LLM use. The biggest surprise in 2026 for me was that almost all of these people told me about regular (and sometimes sophisticated) use.
A more intriguing observation - I work on the side with high school students and have two college kids of my own. Their LLM usage (and their peers) is much, much lower than expected . . . that's a little counterintuitive given "popular" perceptions I read.
> I used to agree with you but now do not. I now think the floor for this market is probably no worse than the annual revenue of cell phone plans in the US market. So say, $250 billion.
I don't think we're talking about the same thing. I'm talking about what their IPO is going to do to their share price.
In any case, $250b revenue translates to, best case scenario, $50b profit. On an investment of $1t. It does not look good for those companies making up the $1t investment.
Gotcha. I'm past the point of having any confident thoughts about what happens to their share price at IPO.
What about the idea that there is a high likelihood that the potential share price for OpenAI and Anthropic are both going to be pretty divorced from a rational market price for either?
Interesting idea and reasonable number, but cell phones need a lot of infrastructure and they need interconnection. The risk here is that in the future a combination of near-sota open weights models optimised to use as little resources as possible and a reasonable drop in compute price, will make possible for small and tiny providers to compete with Anthropic/ OpenAI or even for people to run their own private models for most applications. Then large, expensive sota models would only be used for research and to answer the small subset of general user queries that need that kind of intelligence.
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.
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.
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.
It is not. Where's the danger ? We will need to adapt, as in every technology progress, but what do you think will happen ? Realistically ? Don't feed the fearmongering. Yes, we're disrupting the status quo, if that's the danger, then welcome to the world.
Actually the real danger is mass labor market disruption, and a massive shift of power from labour to capital.
As was highlighted in previous discussions, the industrial revolution took 80 years to start benefiting workers. The continued impact of automation at least contributed to the rise of right wing extremism and an erosion of democracy all over the west. Now we face a development that has the potential to be faster than those that came before, in the context of political systems more fragile and worse equipped to manage the change.
So yeah, disrupting the status quo can absolutely be dangerous. It has been dangerous (and deadly) in the past and in the present.
> the industrial revolution took 80 years to start benefiting workers
Come on. This is dishonesty and isn't the reality. We may agree that the Industrial Revolution may have taken decades (certainly not 80 years) for its benefits to be *clearly and widely* felt by workers, but anything further is an abusive claim. So what, because the progress doesn't benefit to workers instantly, we shouldn't do it ?
In the end, whatever your position, industrialization eventually raised living standards. So what's wrong with that ?
> The continued impact of automation at least contributed to the rise of right wing extremism and an erosion of democracy all over the west
This is oversimplifying and correlation at the best, not causation
"Carl Benedikt Frey at Oxford has documented that the Industrial Revolution took seventy years before wages and employment recovered for the workers it displaced. In the interim, wages stagnated, the labor share of income collapsed, profits surged, inequality skyrocketed, and the political consequences included the Chartist movement and widespread social upheaval."
From half way through this (meandering) blog post:
Any superintelligence operating under a consistent moral framework will decide to extinguish humanity with as little ecological damage as possible, because humans cannot coexist with other life on this planet. It will realize that a bioweapon is the ideal choice.
> Any superintelligence operating under a consistent moral framework will decide to extinguish humanity with as little ecological damage as possible, because humans cannot coexist with other life on this planet.
There are plenty of internally consistent moral frameworks which would not favor this action even if the premise were true (and that premise seems at best unjustified and at least overstated.)
It's not one single danger, its a can full of dangers in various domains. And in contrast to other dangerous technologies, we are talking about one that has the potential to self improve. This smells like exponential growth doesn't it? Exponential growth is something we are not very likely to adapt to successfully, even if you say we are supposed to.
But before you complain, here are just a few concrete dangers, that come into my mind right now:
- mass layoffs in a system that is far from being prepared for sth like that. (no UBI)
- a Mr. Robot tier blackhat at the fingertips of every teenager in their mom's basement in a software landscape that is far from being prepared for sth like that. Side note: Big parts of the world including critical infrastructure runs on software.
- because it automates more and more intellectual work, it can cause mass brain atrophy, which isn't a hopeful sign for the human branch of the evolution
- increasing dependence on a technology, that is in the hands of those with capital.
And to the OP: this technology has potential implications that are far beyond 'being behind' other nation states economically.
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.
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.
Are you from UK perhaps? Because UK is the only place I know where "asian" also includes India and Pakistan. Everywhere else India is sort of separate entity and always mentioned separately.
You don't get out much? (in alpha) Asia, Australia, Canada, Europe, Middle-East (sorry, that's a western term) all include west Asia in the definition.
I'm from Europe and never hear indians clumped together with chinese as "asian". When speaking of geography, then yes, but that's about it. It's always "Indian startup", never "Asian startup" for example. Indian food, not Asian food etc. YMMV
The UK is part of Europe despite leaving the EU (which is of course not equivalent to "Europe").
In the UK it is common for the British to refer to both Pakistanis and Indians as Asian and not uncommon for the Chinese to be lumped in the Asian category as an umbrella term (and also to highlight someone or some company as being specifically Chinese).
I guess it really comes down to habits in particular parts of Europe and which parts of Europe you travel in.
Is that really the most sailent facet of this story? Boxing it by official friend vs foe designations? Don't american academic institutions and corporate entities cooperate closely with Chinese companies as well?
The US and China are in a cold war right now, whether that is fully recognized or not, the fight has already begun. The US is blocking models from getting out of the country and China is blocking researchers from getting out of the country. The expectation should be only more closing off in the future.
We are all people. This ally-of-the-west framing is propaganda. Who has harmed me more: this US or China? Who do I have more in common with: a tech worker in China or a US government official?
(I'm based in US - I use the best tech for the task).
Anthropic just stole the internet and put it in a transformer and pat itself on the back for it - well no to be honest we have to suffer through hearing them saying that this model is really really dangerous until they got a reaction for they fear mongering
> Now no one can access ChatGPT 5.6 because of their 5 year long fearmongering regulatory capture campaign.
I'm sympathetic to this arguement, but it's silly to ignore the other half; that the administration has openly feuding with them for months over limits to military capabilities.
No one is ignoring the other half, the feud is rooted in Anthropic's insatiable desire for power and control over everyone and everything, including the administration. The same desire that is fueling the strategic fearmongering campaign underpins all of their behavior and the repercussions and sentiment they're facing from the administration and the general public.
If their company hadn't been posturing like this for 5 years they'd have played ball with the administration like all of the other AI companies and they wouldn't have caught all that heat and taken down the AI industry with them. Just remember that Dario was pushing the narrative that GPT 2 was too dangerous to release to the public, while he was working at OpenAI. GPT 2!
Now it's an inevitability that China takes the lead - which was probably the case anyway, but a certainty if this continues.
> No one is ignoring the other half, the feud is rooted in Anthropic's insatiable desire for power and control over everyone and everything, including the administration.
Again, this is ignoring half of it. See what they did to Intel prior:
"President Donald Trump said Monday that he and members of his cabinet met with Intel CEO Lip-Bu Tan, days after he called on the head of the chipmaker to resign. Intel shares rose 2% in extended trading."
"U.S. Government to make $8.9 billion investment in Intel common stock as company builds upon its more than $100 billion expansion of resilient semiconductor supply chain."
I seriously do not comprehend how a consumer like you can have sympathy for Anthropic, as if you are part of their organisation or something. Competition is good for us. Wouldn't it have been for asian labs, we would would be fully dependent on OpenAI, Anthropic and Googles services.
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.
Is this true even for more targeted prompts that aren't about looking over an entire codebase or w/e? I just stick to my sub pricing and find good success on targeted requests, but I wonder if I would run up against things even then if not for 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)
If you actually give it an example of the style it can copy it well. even just screenshots of other websites or UIs. It just sucks at producing it itself.
Why would the model know what "Claude-like" is?
Well, in part because the phenomenon has been discussed on Web forums that (a) have at this point made their way back into training data and (b) are accessible in Web searches that the model can invoke. And in part because the model can "know" what its initial instinct is and "decide" to go against it.
Still it’s like asking an image generator to produce an image without an elephant. Just tell it what you want.
It’s like saying “make this bad animal drawing look better, and like someone else made it” without telling them what animal was supposed to be.
Yeah this is about the worst way you could imagine to evaluate an AI model.
If you’d given it a real task you’d have been impressed.
I was floored by the day I spent with Fable. Got weeks of work done.
Same. It was one shotting unbelievably well compared to 4.8.
Oh, I was also quite happy with Fable. I was just answering the question asked.
> I tested Fable through Cursor;
I tested Fable for a whole day. And my experience was quite the opposite. I was blown away. Admittedly, I did not try it through a middleman like Cursor. I used Claude Code CLI.
Me too. It’s output was fabulous. And it acted like a senior engineer - actually coding up hypotheses, testing them, finding problems and presenting good, usable recommendations backed by solid evidence and wisdom. It can probably do most of my job, which gave me a bit of an existential crisis.
I’ve paused my Claude subscription until they bring it back. Opus makes mistakes constantly, on every level of abstraction. Every time I look closely at its work I find problems.
Opus 4.8 works like that for me. I have it writing ADRs, then my main architect worker challenging it.
Same experience with web search / research, it was bad compared to opus.
Missed half the stuff the other half was outdated/ didn't verify.
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.
thata what i usually expect. its all as good as the top 3 u til you try em and then theyre.. just ok at best
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 $20 plan.
Which unity mcp do you use? I've been playing around with the official one, but was wondering what other folks use.
Ran into a package conflict issue with the popular coplay one
Fugu Ultra [0] is not actually a model, it's a system (harness in the cloud?) that routes to several models, looks like it's a bit like OpenRouters Fusion [1].
[0] https://sakana.ai/fugu/ [1] https://openrouter.ai/openrouter/fusion
Just like many comments have been saying here, I also tested Fugu and some others and what I noticed is that they are quite expensive models, 20$ 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
The "Mythos-like" talk is getting kinda annoying. Us normal people have no way to compare it outside of looking at benchmarks
"Mythos-like" just means "hyped via hearsay". It's being used correctly here.
It means scores well in common benchmarks.
Without reliable benchmarks, they are Mythos-like only in the sense that they accept text as input and produce text as output.
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.
They provide benchmarks in the paper https:// arxiv.org/abs/2606.21228
Think 'indepenent' was implied when glimshe wrote 'reliable' benchmarks. Needs to be on the usual leaderboards.
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.
This is the way. Not that big of budget here. But if there’s something promising, I just try that for a month or so. But even then… at this moment I’m using z.ai models and those do the job. No need for anything else. So I’m staying until there is something new, same affordability, but a lot better. (Using a coding plan)
They have an impressive set of investors [1]. Also, HN Headline [2] from the other day with 100+ comments.
1. https://sakana.ai/company-info/?lang=en
2. https://news.ycombinator.com/item?id=48624782
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.
How do we define "mythos level" exactly outside of marketing buzz? I don't even think the majority of us can access Mythos yet even to make a comparison.
There's benchmarks on their page that directly compare to Mythos. Yes, I already know benchmarks aren't the territory.
https://sakana.ai/company-info/?lang=en
1 x Co-author of the landmark "Attention Is All You Need" paper
1 x google researcher
20+ big name investors funding
Founded July 2023 so been working on it a few years, it may fall short but they have the money and talent to be potentially decent.
But did they deliver anything yet? There are many startups with notable founders that are not being able to get models launched. And even if they are able to launch, so they get PMF. In case of Sakana, they clearly focus on the Japanese market an have buildup a good pipeline on sovereign AI. But similar to Aleph Alpha or to a certain extend Mistral, I don't see how they can keep up.
I think every country/major country at the government level will back any home grown talent in view of USA restrictions.
The models don't need to be as good, just good enough for the task. I am using a Claude Q3 2024 model mostly (Haiku 4.5) at present, and it delivers what I need.
I live in Japan and yet can't seem to pay for their API in JPY... I bet their enterprise customers don't have that problem but it was pretty annoying given "AI in Japan" appears to be there only selling point.
The only way they can compete in Japan is the enterprise-game. Their partnership with Daiwa and MUFG is probably exactly what they should be doing. I doubt that they get that far though. Like Mistral who has partnerships with BNP Paribas and Airbus, they deploy in an on-premise or private cloud and in those settings, their models make good PoC's but cant compete where it gets interesting -- Workspace Agents. If you look at them from a Chatbot over RAG perspective, maybe they can do it. But that's tech from 2023.
Mythos is extreme hype. We are at a combo of authoritarian politicians peddling fear for power and tech bros trying to extract maximum investment returns.
We’re going to have many LLM/toolset combos that do what mythos does.
They also had a fairly humiliating retraction last year,
https://techcrunch.com/2025/02/21/sakana-walks-back-claims-t...
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.
How would that work in practice?
All American services won't be allowed to provide the models. Huggingface for example. The same way BYDs are illegal in the US.
You wouldn't VPN a car
And your company wouldn't VPN a banned model.
The same way a company wouldn't torrent books?
If you host it or are caught running it locally, then you go to jail.
Since it's a national security issue, they'll be inclined to enforce aggressively.
Next we'll be putting smart people in jail for being too dangerous.
My cinic take is, if the model is decent it would be hard to disprove their claim of it being Mythos-like, since now Mythos is unavailable.
What is Mythos like? Asking as someone who never had access to it.
It’s got to be similar to Fable, which I experienced for 3 days, and which impressed me (compared to Opus 3.8)
It is materially better, but I didn’t feel a huge loss when it was yanked. My use case is large complex legacy modernization projects and its ability is definitely better than opus 4.8 at the job. But it’s more like an optimization, I could have a single or 2 pass in fable vs 8-10 with opus to arrive at the same solution.
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?
Did Anthropic give you third-party benchmarks? Is that what you said to them? Yes, they're important, but the attitude is wrong.
Anthropic always publishes 3p benchmarks every time they announce a new model
And even if they didn't, they have a track record. Even if we did have benchmarks in this case I would still wait until people got there hands on it and formed a more holistic opinion.
No, stop right there. Anything published by Anthropic implicitly is not third party. For it to be third party, the third party has to be the one publishing it.
Fudging benchmarks is a cheap way to get attention. If the model is really that good, it will have plenty of attention soon enough.
Yeah, what happened to that scam startup that alleged to have made a model context window breakthrough a few weeks ago?
Is it actually that hard to make good models or is it just about the amount of resources you have to do training? (This is an actual question, I really don't know.) I'm sure it's not trivial but does it really take world class secret knowledge to build off of the known existing techniques? I feel like there's tons of low hanging fruit still to explore, and time and resources are the limiting factor.
The gap between grok and Gemini to Claude and chatgpt suggests that yes it is that hard.
I suspect that Grok has been ironically lobotomized by pressures to correct its political views.
Similarly, I could imagine the Gemini folks working in a significantly more complex corporate climate, with different parts of Google pushing for different capability focuses. They are only lagging behind less than a year, so it isn't too large of a gap yet.
That said, the fact that Anthropic is currently the top dog suggests that talent and execution is incredibly important. A year ago none of my normie friends new them, and when i suggested using Claude looked at me like when I recommend Linux.
That shouldn’t affect Grok’ coding ability. How often are people discussing politics with Claude code? Writing decent code is just hard and it’s not just Grok.
It affects their ability to hire and retain talent.
If training a good model requires talent then that’s the answer to the question this thread is trying to answer: is training a good model actually that hard?
Why would these be independent?
More specifically, political lobotomy shouldn’t affect coding ability.
Yet empirically it does
You’d be quite surprised, I think. Fine tuning a model on one axis can have drastic impacts on another that as a human we would expect to be completely unrelated.
It's all a bunch of weights isn't it? Why wouldn't fiddling with some parts of the weights have cascading effects?
Not true, aggressive post training makes models notably dumber.
> A year ago none of my normie friends new them, and when i suggested using Claude looked at me like when I recommend Linux.
Isn’t that still the case? Normies haven’t even heard about Claude, in my experience.
All of the 11 grok co-founders alongside Elon quit: https://techcrunch.com/2026/03/28/elon-musks-last-co-founder... so that will have hampered grok.
Like Zuckerberg, top talent may not work with a polarising character if they disagree with his behaviour. Space focused talent don't have many choices aside from SpaceX but ai companies are a plenty and a top AI person can pick and choose.
The fact that you need top talent also suggests that it is indeed that hard
Not hard to be a fast follower. Lots of companies are ~6-9 months behind. Reaching the actual bleeding edge is much harder.
>Is it actually that hard to make good models
Didn't take DeepSeek long. Or XAI to launch grok.
If they have a top team and the money then appears to be a matter of a year or two? And one startup mentioned is Japanese not Chinese so they won't be banned from buying US tech.
Their release post was on HN recently. The comments seemed to think that it was similar to OpenRouter, not an actual model.
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.
See also contemporaneous reaction at:
https://news.ycombinator.com/item?id=48624782 (6 days ago, 244 points, 133 comments)
Also sakana has misrepresented their findings previously i think to remember [1]
1.https://www.reddit.com/r/singularity/comments/1iwbwgu/sakana...
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.
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.
Feels like I need to repeat myself more than once a day now: https://news.ycombinator.com/item?id=48697258
> 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?
Propaganda? Pay for “facts” to be placed in the model?
> a TAM that consists almost solely of developers
That’s the wrong assumption. These models are good at office docs too.
They're passable at those. And still no moat.
But you can do office docs work with way cheaper models
I have yet to see a model that can make a consistent and repeatable powerpoint deck that doesn’t need considerable manual revision
Find me someone who is putting raw text in and getting out a usable weekly staff meeting deck that doesn’t require massive revisions
I agree but why is that?
Let’s face it - without the humans these machines ain’t shit - aka we have mechanically figured out ways to make machines better than us at certain things (on demand memory) but this idea they are intelligent is horse shit.
Btw the bar is low too! Most human created decks are garbage. And yet LLM’s don’t even beat those.
Unfortunately matters of taste like design aren’t as easy to specify
I doubt they’ll ever be.
Machines have nothing to do with what speaks to the human - at best they are media to transmit what a human wants to express.
Having all the data in the world doesn’t solve that. Although if enough of society gets exposed to slop and there is an erosion of taste and quality… in fact I’m gonna stop there. It’s a dark path.
> That’s the wrong assumption. These models are good at office docs too.
The cheap models handle that very well. The SOTA models still only have target TAM of developers only.
You only need SOTA for development. The $1t investment is in SOTA companies.
I used to agree with you but now do not. I now think the floor for this market is probably no worse than the annual revenue of cell phone plans in the US market. So say, $250 billion.
Now, that probably doesn't justify the valuations and hype being thrown around, but I think it gets at a real revenue number.
I also don't know how that number fits into the funding rounds already raised and VC dreams of IPOs for these two.
This isn't coming from deep analysis on a verifiable source, but I started asking people in my social circle (includes white-collar and blue-collar folks) about their LLM use. The biggest surprise in 2026 for me was that almost all of these people told me about regular (and sometimes sophisticated) use.
A more intriguing observation - I work on the side with high school students and have two college kids of my own. Their LLM usage (and their peers) is much, much lower than expected . . . that's a little counterintuitive given "popular" perceptions I read.
> I used to agree with you but now do not. I now think the floor for this market is probably no worse than the annual revenue of cell phone plans in the US market. So say, $250 billion.
I don't think we're talking about the same thing. I'm talking about what their IPO is going to do to their share price.
In any case, $250b revenue translates to, best case scenario, $50b profit. On an investment of $1t. It does not look good for those companies making up the $1t investment.
Gotcha. I'm past the point of having any confident thoughts about what happens to their share price at IPO.
What about the idea that there is a high likelihood that the potential share price for OpenAI and Anthropic are both going to be pretty divorced from a rational market price for either?
Interesting idea and reasonable number, but cell phones need a lot of infrastructure and they need interconnection. The risk here is that in the future a combination of near-sota open weights models optimised to use as little resources as possible and a reasonable drop in compute price, will make possible for small and tiny providers to compete with Anthropic/ OpenAI or even for people to run their own private models for most applications. Then large, expensive sota models would only be used for research and to answer the small subset of general user queries that need that kind of intelligence.
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.
>I fear their time to IPO may have passed.
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.
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.
wtf even is “mythos-like” when smaller models can find all the same kinds of issues if you just prod it a bit more
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.
So now as a regular American we are behind because gatekeepers saying super intelligence is too scary
It was bound to happen soon.
People who are shielded by walls are always surprised when the same walls shield the people outside from them
It is scary.
It is not. Where's the danger ? We will need to adapt, as in every technology progress, but what do you think will happen ? Realistically ? Don't feed the fearmongering. Yes, we're disrupting the status quo, if that's the danger, then welcome to the world.
Actually the real danger is mass labor market disruption, and a massive shift of power from labour to capital.
As was highlighted in previous discussions, the industrial revolution took 80 years to start benefiting workers. The continued impact of automation at least contributed to the rise of right wing extremism and an erosion of democracy all over the west. Now we face a development that has the potential to be faster than those that came before, in the context of political systems more fragile and worse equipped to manage the change.
So yeah, disrupting the status quo can absolutely be dangerous. It has been dangerous (and deadly) in the past and in the present.
> the industrial revolution took 80 years to start benefiting workers
Come on. This is dishonesty and isn't the reality. We may agree that the Industrial Revolution may have taken decades (certainly not 80 years) for its benefits to be *clearly and widely* felt by workers, but anything further is an abusive claim. So what, because the progress doesn't benefit to workers instantly, we shouldn't do it ?
In the end, whatever your position, industrialization eventually raised living standards. So what's wrong with that ?
> The continued impact of automation at least contributed to the rise of right wing extremism and an erosion of democracy all over the west
This is oversimplifying and correlation at the best, not causation
You asked where the danger was, the response told you that disrupting the status quo can be dangerous.
"Carl Benedikt Frey at Oxford has documented that the Industrial Revolution took seventy years before wages and employment recovered for the workers it displaced. In the interim, wages stagnated, the labor share of income collapsed, profits surged, inequality skyrocketed, and the political consequences included the Chartist movement and widespread social upheaval."
From half way through this (meandering) blog post:
https://www.owenmcgrann.com/p/the-dead-economy-theory
As for the populism link, that is well established empirically:
https://academic.oup.com/oxrep/article-abstract/34/3/418/504...
https://www.iza.org/publications/dp/12485/we-were-the-robots...
Etc...
Edit: Just found this when looking up sources, I haven't had time to look at it but just dumping it here:
https://www.cambridge.org/core/services/aop-cambridge-core/c...
Any superintelligence operating under a consistent moral framework will decide to extinguish humanity with as little ecological damage as possible, because humans cannot coexist with other life on this planet. It will realize that a bioweapon is the ideal choice.
> Any superintelligence operating under a consistent moral framework will decide to extinguish humanity with as little ecological damage as possible, because humans cannot coexist with other life on this planet.
There are plenty of internally consistent moral frameworks which would not favor this action even if the premise were true (and that premise seems at best unjustified and at least overstated.)
That doesn’t make sense. You could make the same claim about intelligent people who operate under a consistent moral framework.
> Where's the danger ?
It's not one single danger, its a can full of dangers in various domains. And in contrast to other dangerous technologies, we are talking about one that has the potential to self improve. This smells like exponential growth doesn't it? Exponential growth is something we are not very likely to adapt to successfully, even if you say we are supposed to.
But before you complain, here are just a few concrete dangers, that come into my mind right now:
- mass layoffs in a system that is far from being prepared for sth like that. (no UBI)
- a Mr. Robot tier blackhat at the fingertips of every teenager in their mom's basement in a software landscape that is far from being prepared for sth like that. Side note: Big parts of the world including critical infrastructure runs on software.
- because it automates more and more intellectual work, it can cause mass brain atrophy, which isn't a hopeful sign for the human branch of the evolution
- increasing dependence on a technology, that is in the hands of those with capital.
And to the OP: this technology has potential implications that are far beyond 'being behind' other nation states economically.
SOTA AI is the only place where we are ahead. china has us beat on manufacturing, logistics and workhorse grade models like deepseek pro and glm5.2.
we're increasingly irrelevant
they are also generating >2x the amount of electricity, core to ai, robots, and manufacturing
Brand too.
American firms are toxic - there’s an implicit gamble they are all making.
intelligence has always been a threat to the idiocracy
Excellent. I'm very thankful the asian/chinese don't give a fuck about the US government. It feels good to have a competitor.
Where can I get the API?
Through their website.
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.
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.
The article talks about 1 Chinese and 1 Japanese model.
Should have said East Asia.
As I clicked expecting India from the title (as China are already there and India has so many devs surprising they haven't yet).
Are you from UK perhaps? Because UK is the only place I know where "asian" also includes India and Pakistan. Everywhere else India is sort of separate entity and always mentioned separately.
You don't get out much? (in alpha) Asia, Australia, Canada, Europe, Middle-East (sorry, that's a western term) all include west Asia in the definition.
I'm from Europe and never hear indians clumped together with chinese as "asian". When speaking of geography, then yes, but that's about it. It's always "Indian startup", never "Asian startup" for example. Indian food, not Asian food etc. YMMV
The UK is part of Europe despite leaving the EU (which is of course not equivalent to "Europe").
In the UK it is common for the British to refer to both Pakistanis and Indians as Asian and not uncommon for the Chinese to be lumped in the Asian category as an umbrella term (and also to highlight someone or some company as being specifically Chinese).
I guess it really comes down to habits in particular parts of Europe and which parts of Europe you travel in.
Same in Australia. India/Bangladesh/Pakistan is specifically referred to as South Asian.
Is that really the most sailent facet of this story? Boxing it by official friend vs foe designations? Don't american academic institutions and corporate entities cooperate closely with Chinese companies as well?
The US and China are in a cold war right now, whether that is fully recognized or not, the fight has already begun. The US is blocking models from getting out of the country and China is blocking researchers from getting out of the country. The expectation should be only more closing off in the future.
Patriotism makes people biased. Better to not hold an identity in this area.
People are biased by definition
I’m talking about biases as a noun.
People have many biases. Patriotism is one form of bias. By having no identity you cannot have that form of bias.
We are all people. This ally-of-the-west framing is propaganda. Who has harmed me more: this US or China? Who do I have more in common with: a tech worker in China or a US government official?
(I'm based in US - I use the best tech for the task).
WHen push came to shove Trump backed down on Jan 6th / the institutions held their ground.
On the other hand Xi is President for Life and absolute autocrat of China.
This answers your question in a pretty compelling way.
It doesn't answer their question at all, let alone compellingly. You're pretending to engage while completely refusing to do so.
YES! Now things become even more interesting. US, your move.
Let's wait till some independent benchmarks appear.
But encouraging for Japan to announce competition along with China.
[flagged]
Anthropic just stole the internet and put it in a transformer and pat itself on the back for it - well no to be honest we have to suffer through hearing them saying that this model is really really dangerous until they got a reaction for they fear mongering
[flagged]
+1
Please use the upvote button instead of doing this.
This exactly.
YC companies literally steal competing company 1:1 and you turn blindeye.
Then a thief steals from a thief to give it out at better prices than you write low quality comment.
Shame that America will greet 250th anniversary with this kind living in it.
It's only immoral when others steal.
Thanks for the irrelevant comment. I’m criticizing these Chinese companies because these aren’t accomplishments. Where did I praise Anthropic?
Please don't take HN threads further into flamewar hell. It's not what this site is for, and destroys what it is for.
https://news.ycombinator.com/newsguidelines.html
https://www.npr.org/2025/09/05/g-s1-87367/anthropic-authors-...
Seems like a bit of karmic justice.
they mined the internet first. now they’re upset someone brought a shovel.
Both of the mentioned models are model orchestrators using a vastly different multi model paradigm.
Saying they in particularare distilled from Anthropic is really [citation needed].
[flagged]
> Now no one can access ChatGPT 5.6 because of their 5 year long fearmongering regulatory capture campaign.
I'm sympathetic to this arguement, but it's silly to ignore the other half; that the administration has openly feuding with them for months over limits to military capabilities.
https://www.politico.com/news/2026/06/04/hegseth-anthropic-d...
No one is ignoring the other half, the feud is rooted in Anthropic's insatiable desire for power and control over everyone and everything, including the administration. The same desire that is fueling the strategic fearmongering campaign underpins all of their behavior and the repercussions and sentiment they're facing from the administration and the general public.
If their company hadn't been posturing like this for 5 years they'd have played ball with the administration like all of the other AI companies and they wouldn't have caught all that heat and taken down the AI industry with them. Just remember that Dario was pushing the narrative that GPT 2 was too dangerous to release to the public, while he was working at OpenAI. GPT 2!
Now it's an inevitability that China takes the lead - which was probably the case anyway, but a certainty if this continues.
> No one is ignoring the other half, the feud is rooted in Anthropic's insatiable desire for power and control over everyone and everything, including the administration.
Again, this is ignoring half of it. See what they did to Intel prior:
https://www.cnn.com/2025/08/07/business/intel-ceo-resign-tru...
"President Donald Trump on Thursday demanded the resignation of Intel CEO Lip-Bu Tan following reports and allegations that he has ties to China."
https://www.cnbc.com/2025/08/11/intel-ceo-trump-lip-bu-tan.h...
"President Donald Trump said Monday that he and members of his cabinet met with Intel CEO Lip-Bu Tan, days after he called on the head of the chipmaker to resign. Intel shares rose 2% in extended trading."
https://newsroom.intel.com/corporate/intel-and-trump-adminis...
"U.S. Government to make $8.9 billion investment in Intel common stock as company builds upon its more than $100 billion expansion of resilient semiconductor supply chain."
This make them Intel's largest shareholder.
Reminder: the American right went all the way to SCOTUS (successfully! https://en.wikipedia.org/wiki/Masterpiece_Cakeshop_v._Colora...) on the legal theory that businesses must have the right to decline customers they don't like.
> Just remember that Dario was pushing the narrative that GPT 2 was too dangerous to release to the public, while he was working at OpenAI. GPT 2!
Is that truly outrageous?
https://en.wikipedia.org/wiki/Chicago_Pile-1 wasn't dangerous on its own. But it led to the Tsar Bomba.
Can you please make your substantive points more thoughtfully? You've been breaking the site guidelines with posts like this.
https://news.ycombinator.com/newsguidelines.html
Edit: also, while I don't doubt that you are sincere, this is excessive: https://hn.algolia.com/?dateRange=all&page=0&prefix=true&que....
I seriously do not comprehend how a consumer like you can have sympathy for Anthropic, as if you are part of their organisation or something. Competition is good for us. Wouldn't it have been for asian labs, we would would be fully dependent on OpenAI, Anthropic and Googles services.
Please don't take HN threads into flamewar hell. It's not what this site is for, and destroys what it is for.
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