Sakana AI adds two models to its multi-agent orchestration line

Sakana AI has released two new models in its Fugu family: Fugu Max and Fugu Ultra v2. Fugu is not a single foundation model. It is a learned orchestrator that routes work across a pool of other models behind one API. The new release tunes that same architecture for two different goals: Fugu Max chases the best output per dollar, while Fugu Ultra v2 aims for peak capability on hard, multi-step tasks.

Both models are live through Sakana's OpenAI-compatible API. There are no open weights for self-hosting, and Sakana does not offer the service in the EU/EEA.

How the orchestrator works

The technical report describes Fugu models as language models in their own right: they read a query and build an agentic scaffold for it on the fly. Training combines large-scale fine-tuning, evolutionary algorithms and reinforcement learning. The system builds on two ICLR 2026 papers. TRINITY uses a lightweight evolved coordinator that assigns Thinker, Worker or Verifier roles across turns. The Conductor is trained with reinforcement learning to discover natural-language coordination strategies and focused prompts.

Sakana frames the design around the Pareto frontier: gaining quality costs more, cutting cost loses quality. Sending a simple data lookup to a multi-trillion-parameter model wastes money, so the company argues for picking the cheapest machinery that can still solve the task.

Where the two differ

Fugu Max widens the pool of models Fugu can orchestrate, adding a large set of open-weights and specialized models, including the NVIDIA Nemotron family through Sakana's collaboration with NVIDIA. Each task goes to the leanest model capable of solving it. Sakana places Fugu Max within striking distance of elite models at 2x to 6x lower cost, citing SWEFish, an internal benchmark built from its own coding challenges. That figure is a vendor signal.

Fugu Ultra v2 targets complex reasoning, autonomous research and full-stack software development, with its largest gains on sustained reasoning over visual and structured data. Fable 5, Fable 5.1 and GPT-6-Astra are not in its agent pool. The training cutoff is 28 August 2026.

Sakana's stated point is frontier output without depending on any single proprietary model, which the research team says reduces exposure to vendor lock-in, API revocations and sudden service cutoffs.

Details are available in the MarkTechPost report.