Edited by Lin Xin
Yesterday, a model was released that claims to be comparable to "Anthropic's Fable 5 and Mythos Preview" — Sakana Fugu (pufferfish), which has already racked up tens of millions of views on X.
Has a new player entered the arena of top-tier AI models?
Not long ago, there was a viral trend on X featuring "LeChaton Fat." So, is Sakana Fugu genuinely that powerful, or is it just marketing hype?
Note: "LeChaton Fat," translating to "Fat Cat" or "Chubby Kitty," isn't a real model. The chatbot from French AI startup Mistral AI was originally called Le Chat. This was later turned into a meme claiming it was a super-model with "30 trillion parameters" that would "crush all its competitors."
According to reports, Fugu Ultra matches Anthropic's Fable 5 and Mythos Preview on the most difficult engineering, science, and reasoning benchmarks, and outperforms Gemini 3.1 Pro, Opus 4.8, and GPT-5.5 on tasks like AutoResearch, mechanical design, and financial forecasting. Crucially, it involves no export control risks.
What is the background of a model that sounds this strong?
Who is Sakana AI?
So, the journalist did some digging into the company. Sakana focuses on developing AI based on nature-inspired principles, using "evolutionary model merging" technology. Founded in July 2023, it became Japan's fastest-growing AI unicorn in less than a year.
The founders are just as impressive: Llion Jones, one of the eight authors of the Transformer paper, and David Ha, a former research scientist at Google Brain and Stability AI.
Llion Jones expressed at the TED AI conference in San Francisco in late October 2025: "I am honestly so done with Transformers."
Is Sakana Fugu an AI Service Wrapper?
The newly released Sakana Fugu is billed as a "multi-agent orchestration system," like an intelligent conductor:
You only call one API endpoint.
Internally, it automatically selects the best models (Claude, GPT, etc.), breaks down tasks, allows multiple models to collaborate, and finally synthesizes an answer.
Fugu Ultra claims performance rivaling frontier models (like Fable 5, Mythos).
Two versions were launched:
Fugu: Balances performance and latency, suitable for everyday coding, code reviews, and interactive scenarios.
Fugu Ultra: Optimized for answer quality, suitable for highly complex problems.
But some technically adept individuals have dug into Fugu's technical report:
Fugu itself is a closed-source scheduler, and its underlying functionality still depends on closed-source large language models. Previously, you at least knew which model you were using; now, you don't even know "which model was used, how many tokens were consumed, or what the cost was." It's a complete black box.
Consequently, some developers are criticizing it: "Isn't this just an AI service wrapper?"
Official Benchmarks Are Impressive:
Matches Fable 5 and Mythos Preview in Benchmarks
In the official report, Sakana Fugu's benchmark performance stands out: It is on par with Fable 5 and Mythos Preview in engineering, science, and reasoning benchmarks.
And in specific scenarios, it outperforms Opus 4.8, GPT-5.5, and Gemini 3.1 Pro.
Beyond the benchmarks above, Fugu performs exceptionally well in lengthy and complex real-world workflows.
In automated data science research: Early users ran Sakana Fugu in a nearly fully autonomous research mode and found it could make significant progress with almost no human intervention. For us, this is exactly what Fugu Ultra was designed for: handling open-ended, multi-step work where the system needs to explore ideas, run experiments, analyze failures, improve methods, and make progress over time.
Real-World Testing: Fugu's Price is its Only Mythical Feature
After Sakana Fugu was released, some fast-moving teams conducted their own tests: Sakana Fugu's performance was unexpectedly close to GLM 5.2's level, but its price was 17 times higher!
We gave the same request to four models: Build a complete real-time trading platform, including front-end and back-end components, fetching real-time market data for eight trading instruments from an external API, with a custom dark-themed user interface.
Output: Fugu Ultra — 22,225t, $0.51
Opus 4.8 — 15,802t, $0.31
GPT-5.5 — 11,474t, $0.26
GLM 5.2 — 13,677t, $0.03
Fugu built the most comprehensive and feature-rich trading platform in this evaluation. GLM 5.2 came in close second, with an equally complete multi-panel interface and real-time data, but at a much lower cost. Opus and GPT also performed well, striking a better balance between quality and cost, and achieved decent results.
In the comments of the post, one netizen stated bluntly: "Fugu is only mythical when it comes to its price."
Netizens: Not as Good as Advertised
Regarding Fugu, the comments from netizens are split into two camps. One side says, "It's not as good as advertised"; the other side says, "It's remarkably outstanding."
"This doesn't match Mythos' performance at all; it's not as good as advertised."
Besides the skepticism, some have given Fugu a high score after using it.
Final Words
The community's attitude towards Sakana Fugu is mixed. Some believe that the multi-model fusion approach can achieve performance superior to any single constituent model, potentially representing a future direction for development; others feel it's an OpenRouter Fusion API clone.
What do you all think of the Fugu model?