Sakana Fugu routes models via one API
Sakana AI launched Fugu, one model that orchestrates other AI models behind a single OpenAI-compatible API, in two tiers with published benchmarks and pricing.
Sakana AI released Fugu on June 22, a model whose job is to run other models. Instead of answering directly, Fugu is trained to call a pool of other systems, including copies of itself, and to plan, delegate, check, and combine the results. You reach all of it through one OpenAI-compatible API, so the orchestration happens behind a single endpoint, not in your code.
It ships in two tiers. Plain Fugu is tuned for everyday work where latency matters, like coding tools, review steps, and chat. Fugu Ultra reaches for a deeper pool of agents on harder jobs such as reproducing a paper or security analysis. Sakana reports Fugu Ultra at 73.7 on SWE-Bench Pro and 93.6 on MRCRv2, and says it rivals frontier systems like Claude Fable 5. Subscriptions start at $20 a month; Fugu Ultra runs $5 per million input tokens and $30 per million output tokens. The approach comes out of two ICLR 2026 papers, Trinity and the Conductor, on teaching a system to assemble its own agents instead of hand-wiring a workflow.
Why this matters
The pitch is that you stop choosing the model. For an operator who already juggles three or four tools depending on the task, handing that decision to a router is genuinely tempting, and one bill is simpler than five.
The catch is the same one as always. A vendor’s own benchmark is a starting point, not a verdict, so treat 73.7 like any performance claim and test it on your own work. An orchestrator that hides which model ran your task is also convenient right up until you need to move off it, the lock-in question worth asking on day one. The wider point: open and frontier models are now close enough that routing between them is a product in itself.
Update, July 24: Sakana shipped Fugu-Ultra v1.1, which it says gains up to 7.9 points over v1.0 across its benchmark suite, with the biggest jumps on ProgramBench and Terminal Bench 2.1, at the same price as v1.0. Sakana claims v1.1 now beats Claude Fable 5 on complex coding and reasoning tasks without including Fable 5 in its own agent pool, since Fable 5 isn’t publicly accessible for Fugu to call. As with the original launch, that comparison is Sakana’s own benchmark, not an independently reproduced one, so weigh it the same way, as a starting point for your own testing rather than a settled result.