What Is Fugu? How Sakana AI's Orchestration Actually Works

Fugu in one sentence

Here's the thing that trips people up: Fugu is not a model. It's an orchestrator. When you send it a request, it doesn't answer you itself. It looks at what you asked, breaks the job into parts, picks the best model from a hidden pool for each part, runs them, and stitches everything into one answer that comes back through a single OpenAI-compatible API call. You never see the machinery. That's the whole idea — and if you want the full pricing picture, our breakdown of Fugu Max vs Ultra pricing covers what each tier costs and who it's for.

An orchestra, not a soloist

The best analogy I know is a restaurant kitchen during the dinner rush. You wouldn't ask the pastry chef to grill a steak. The head chef looks at the order, fires the grill station, the sauce station, and the pastry station in parallel, and plates one coherent dish. Fugu is the head chef. The "models" are the stations. A hard question gets split across specialists — one might reason through the plan, another writes the code, a third checks the result — and you get back one clean answer.

This is why Sakana calls it orchestration rather than routing. A router — think LiteLLM or OpenRouter — picks one model for your request based on rules you write, and that's it. Fugu makes the decision itself, model by model, part by part, and it combines the outputs. It doesn't even tell you which models it used. You trade visibility and control for a system that, on paper, beats any single model in the pool on tasks that reward combining approaches.

The research behind it: TRINITY and Conductor

Fugu didn't come out of nowhere — it grows directly out of two papers Sakana published at ICLR 2026, and both read like pure Sakana: coordination treated as something you can evolve and learn, not hand-design.

TRINITY describes an evolved LLM coordinator. Instead of a human writing the rules for who does what, the coordinator itself was shaped by evolutionary algorithms — the same family of techniques behind Sakana's earlier model-merging work. In each turn it assigns roles across the pool: Thinkers plan, Workers execute, Verifiers check. It's evolution applied to teamwork.

Conductor goes one level deeper. It uses reinforcement learning to discover coordination strategies expressed in natural language — the system learns, through trial and reward, which ways of asking models to collaborate actually produce better answers. Not a fixed playbook, but strategies the system found on its own.

Together they answer the obvious objection: "won't a committee of models just argue?" The whole point of the training — large-scale fine-tuning plus evolution plus RL — is teaching the coordinator to run the room.

The timeline so far

Fugu's rollout has moved fast. It entered beta in April 2026, reached general availability in June, and added Fugu-Cyber plus a Claude Code interface in July. Then on September 11, 2026, Sakana launched Fugu Max v1.0 and Fugu Ultra v2.0 — the two flavors the product line now revolves around. Max is the lean one: best output per dollar, built more heavily on open-weight and specialist models including NVIDIA's Nemotron family. Ultra v2 is the heavyweight: highest capability on hard, multi-step tasks.

Does it actually work?

On Sakana's own reported benchmarks, the numbers are striking — with the usual caveat that these are the vendor's numbers, not independent tests. Fugu Ultra v2 scores 48.3 on Chartography against Claude Opus 5's 27.3, and 74.3 on DeepSWE. It claims best or joint-best on five of eight benchmarks. Fugu Max claims the best overall score on six evaluations including Terminal Bench 2.1. The detail that makes these claims interesting rather than just marketing: Ultra v2 posts them without GPT-6 Astra, Claude Fable 5, or Fable 5.1 in its agent pool. It's beating frontier scores using mostly open and specialist models — which is exactly the bet Sakana has been making since the lab was founded.

Why this fits Sakana like a glove

Step back and Fugu looks like the inevitable product of everything this lab believes. Evolutionary model merging treated open models as a gene pool. The AI Scientist treated research as a multi-agent process. Fugu treats inference itself as a coordination problem. Same philosophy, one rung higher. And there's a bigger story underneath the engineering — why orchestration over open models matters for vendor lock-in and even export controls. We dig into that bigger picture here, because the technology is only half of what's interesting.