The big AI labs are locked in an arms race over ever-bigger models. A small team in Tokyo just opted out of that race entirely — and still beats the giants on one benchmark. The trick: it didn't build a new flagship model at all. It built a conductor.
What Fugu Ultra actually is
Fugu Ultra v2, Sakana AI's new release, isn't itself a single large language model. Think of it more as a switchboard: you send a request to one API, and Fugu decides behind the scenes which of several existing models in a fixed pool should handle it — and can even recursively call itself to stitch results back together. It was trained with a mix of standard fine-tuning, evolutionary algorithms, and reinforcement learning (the AI gets rewarded for good answers, roughly like training a pet).
That's not a one-off gimmick — it's Sakana's signature move. The Tokyo lab has spent years combining existing open models with evolutionary algorithms into new, better ones — breeding models instead of expensively training new ones from scratch. Co-founder Llion Jones is one of the eight authors of the famous 2017 "Attention Is All You Need" paper — the work that introduced the Transformer architecture behind ChatGPT, Claude, and practically every modern language model. Apparently, once you've helped lay the foundation for the whole AI wave, you'd rather build conductors than more soloists.
The number that raises eyebrows
On a benchmark called Chartography, which tests visual and tabular reasoning, Fugu Ultra v2 reportedly scored 48.3 — well ahead of Anthropic's Opus 5 (27.3) and Fable 5 (29.5). Notably, neither competing model is even part of Fugu's own model pool.
One caveat, so this doesn't get overblown: that's a single benchmark, not proof of overall superiority. But it does show that clever task-splitting across existing models can keep pace with the industry's biggest names — without owning a frontier model of your own.
What it costs — and what it doesn't
Nothing here is free. As of now, Fugu Ultra v2 costs roughly $5 per million input tokens and $30 per million output tokens, with a 1-million-token context window. Its cheaper sibling, Fugu Max, is aimed at use cases where per-request cost matters more than the top benchmark score, and reportedly costs only a fraction of that. There's no permanent free tier here like ChatGPT's or Claude's — if you seriously test it, you pay per token.
Trying it out — but mostly as a developer
Sakana doesn't offer its own chat app like ChatGPT or Claude. You get access through Sakana's own API or through OpenRouter — the same model marketplace we've covered here before. Signing up there is free and requires no credit card, and you can try Fugu Ultra v2 directly in the Chat Playground. You'll still pay to use it, though: unlike OpenRouter's clearly marked free models, Fugu isn't one of them — it runs against your loaded balance.
Honestly, this isn't a tool most beginners will open after work just to poke around. Fugu Ultra v2 is an API product for developers building their own applications — not a replacement for your everyday chat assistant. But the idea behind it is worth understanding even if you never write a line of code: instead of stuffing one mega-model with ever more parameters, more and more providers are combining smaller, specialized models cleverly. Mixture of Experts already does this inside a single model — Fugu takes the same idea one level up, across entire models.
Maybe that's the real story here: the next big AI leap doesn't have to come from an even bigger model. Sometimes a clever conductor is enough.