Sakana Fugu: A Multi-Agent Orchestration System as a Foundation Model

Sakana AI has a new commercial product, and it is not another giant foundation model trying to out-benchmark everyone else. It is called Fugu, and the pitch is almost contrarian for this moment in AI: rather than scaling one model bigger, Sakana trained a small language model whose job is to call other frontier models, assign them roles, and stitch their outputs into something better than any single one could produce alone.

The company frames this as the throughline of everything it has worked on, from evolutionary model merging to The AI Scientist to ShinkaEvolve to AB-MCTS. Each of those projects, in Sakana's telling, pointed the same way — that coordinated pools of specialized agents beat monolithic scaling. Fugu is described as the commercial product form of that research bet, built on two ICLR 2026 papers, Trinity and Conductor, with what Sakana says are substantial improvements layered on top for performance and usability.

The practical problem Fugu is aiming at is one anyone juggling GPT, Gemini and Claude API keys will recognize. Different models are good at different things, sometimes down to the level of individual problems rather than broad categories, and manually routing tasks between them is both expensive and hard to get right by hand. Fugu is pitched as dropping into that same OpenAI-format API setup with minimal changes, but handling the routing, role assignment and task dispatch itself, adaptively, rather than following a fixed workflow a human designed.

There are two flavors on offer: Fugu Mini, built for lower latency, and Fugu Ultra, the fuller orchestration system aimed at harder, more demanding tasks. Sakana says Fugu has already been running internally as a tool for its own researchers and engineers, and the beta invite is aimed specifically at people testing it in coding assistants like OpenCode and Codex, or in their own engineering and business projects, with Sakana explicitly asking testers to surface where it falls short and what a system like this is actually missing.

The more unusual detail is what happens when Fugu is allowed to call itself. During training it can learn to feed its own prior output back in as context and decide whether to revise its own coordination strategy — a recursive test-time scaling trick where the depth of that self-calling becomes a compute dial you can turn at inference time, no retraining required. Sakana's framing is that a small model, by rereading itself, can inch toward answers that neither it nor any of the individual models it commands could hit in one pass.