Sakana AI launches Fugu, a multi-agent orchestrator delivered as a single model

Sakana AI launches Fugu, a single-model multi-agent orchestrator managing an expert model pool via an API to bypass export controls.

Sakana AI is launching Sakana Fugu, a multi-agent orchestration system delivered as a single model, accessible via a single OpenAI-compatible API. Fugu is itself an LLM, trained to call other models within an agent pool, including recursive instances of itself, to dynamically orchestrate the best models for long, multi-step tasks.

The principle: a request arrives at a single endpoint, and Fugu decides how to process it, alone when sufficient, or by coordinating a team of expert models when the task requires it. Model selection, delegation, verification, and synthesis are managed internally, so that multi-agent complexity never reaches the code. Two variants coexist: Fugu, which prioritizes a balance between performance and latency and integrates with tools like Codex, and Fugu Ultra, tailored for maximum quality on challenging problems, with a deeper pool, for AI research, cybersecurity, or patent investigation. Agents can be excluded from the pool for compliance reasons.

Sakana champions orchestration as the next frontier, after the race for ever-larger models, and primarily makes it an argument for sovereignty: depending on a single provider would be a vulnerability. To this end, it cites recent export controls affecting cutting-edge models, including those from Anthropic, capable of cutting off access overnight, whereas an interchangeable pool would circumvent the cutoff. The company claims frontier-level performance for Fugu Ultra. The product is available now, by subscription or pay-per-use.