Sakana AI's Fugu Turns Model Orchestration Into a Product

Tokyo research lab Sakana AI on June 22 made Fugu generally available, a system it pitches not as another chatbot but as a foundation model whose entire job is to command other models. Where most labs ship a single set of weights and ask you to prompt it well, Fugu exposes one endpoint that decides — request by request — whether to answer directly or convene a pool of specialist models, delegate the work, check the results, and stitch them back into a single reply.

The distinction Sakana keeps drawing is that Fugu is not a router with hand-written if/else rules. It is a language model trained to orchestrate, including recursive calls to instances of itself. The behavior — when to delegate, how the sub-agents should talk to one another, and how to fuse their outputs into something trustworthy — is learned rather than scripted. That design rests on two papers the lab is presenting at ICLR 2026: TRINITY, which evolves a lightweight coordinator that hands out Thinker, Worker, and Verifier roles, and the Conductor, trained with reinforcement learning to discover coordination strategies expressed in plain natural language.

There are two tiers. Plain Fugu is tuned for low latency and is aimed at interactive work — coding assistants, review tools, and chat services. Fugu Ultra trades speed for depth, marshalling a larger pool of expert agents for harder, longer jobs such as paper reproduction, cybersecurity assessment, patent and literature investigation, and automated data-science research. Sakana says Fugu Ultra holds its own against frontier systems including Anthropic's Fable 5 and Opus 4.8, Google's Gemini 3.1 Pro, and OpenAI's GPT-5.5 across engineering, scientific, and reasoning benchmarks. One of the roughly 500 beta testers, quoted in the launch post, called Ultra "significantly better than GPT-5.5," saying it "finds bugs others miss."

Both tiers ship behind a single OpenAI-compatible API, so teams can point existing tooling at console.sakana.ai without rewriting their integration. Sakana is offering subscription tiers for standard use alongside pay-as-you-go pricing for heavier enterprise workloads. The notable caveat: Fugu is not available in the EU or EEA at launch, a restriction the company has not fully explained.

The strategic argument underneath the product is about dependency. Because Fugu's agent pool is swappable, Sakana frames it as a hedge against being locked into any one lab — and, pointedly, as a way to keep working even as export controls and supply constraints reshape who can access which frontier models. The lab calls this "AI sovereignty," and the framing lands in a market where buyers increasingly worry that betting an entire stack on a single vendor is a risk in itself.

Fugu also crystallizes a broader shift the industry has been circling for months: orchestration is becoming the product, not the plumbing. If the hardest engineering problem is no longer building one more capable model but reliably combining the ones that already exist, then the company that owns the coordination layer owns a valuable piece of real estate — regardless of whose weights are doing the work underneath.