Sakana's Fugu Is a Single Model That Commands Many

Most AI labs spend their effort building one model that does everything. Sakana AI has gone the other way. On June 22 the Tokyo company released Fugu, a system that does not try to answer your question at all. It works out which of the world's other models should answer it, hands the work over, checks the result, and gives you back the answer through a single interface.

The pitch is in the name. Fugu, the pufferfish, is harmless across most of its body and lethal in one organ if prepared badly, and Sakana has a habit of naming things after the sea. What the product does is orchestration. It is a language model trained to call other language models, sometimes including copies of itself, and to manage the whole chain of delegation, checking and synthesis without the user seeing any of the plumbing. You send one request to one API. Behind it, a pool of frontier models gets put to work.

One conductor, many players

Sakana ships Fugu in two sizes. The standard Fugu keeps latency low and is aimed at everyday work, coding and code review. Fugu Ultra coordinates a deeper bench of expert agents and is built for hard, multi-step problems where getting the right answer matters more than getting a fast one. On Sakana's own benchmarks, Fugu Ultra holds its own against frontier systems such as Anthropic's Fable 5 and the Mythos preview across engineering, science and reasoning tests, with the company also reporting strong results in automated research, mechanical design, chess and financial forecasting.

Independent benchmarks are not out yet, so those numbers deserve the caution that comes with any vendor's launch-day figures. One early user, quoted by VentureBeat, said that for code review Fugu Ultra was noticeably better than GPT-5.5 and surfaced more than twenty issues a single model had missed. The approach builds on Sakana's published research, including its ICLR 2026 papers Trinity and Conductor, so the orchestration idea is not something dreamed up for a press release.

A hedge against the kill switch

The timing is hard to miss. Sakana is openly selling Fugu as insurance against the kind of disruption the industry has just lived through. When Washington ordered Anthropic to pull Fable 5 and Mythos 5 for foreign nationals earlier this month, anyone who had built on those models lost access overnight. Sakana's argument is that if your application talks to Fugu rather than to any single provider, a sudden ban or outage somewhere becomes a routing problem instead of a crisis. As the company puts it, if one provider restricts access, Fugu reroutes around the disruption.

Pricing reflects the premium. Fugu Ultra runs at roughly $5 per million input tokens and $30 per million output tokens on a pay-as-you-go basis, with subscription tiers around $20, $100 and $200 a month. You are paying for the coordination layer on top of whatever the underlying models cost, which is the catch worth weighing. Orchestration is not free, and a model whose job is to call other models inherits the bills and the lag of all of them.

Still, there is something clarifying about a company betting that the winning move is not to own the smartest model but to be the layer that decides which smart model to use. If that bet pays off, the value in AI shifts from the models themselves to the thing standing above them. Sakana has decided to build that thing first.