Sakana Fugu sells orchestration as the model

Sakana AI launched Fugu on June 22, and it's the most contrarian release of the month. Everyone else is racing to train one bigger model. Sakana trained a model whose entire job is to command other models. Fugu is a multi-agent orchestration system that ships as a single foundation model behind one OpenAI-compatible endpoint. You call one API, and behind it a trained delegator dynamically routes your prompt across a pool of LLMs, including recursive copies of itself, then synthesizes the answer.

The tagline is 'One Model to Command Them All,' and the pitch is exactly the counterintuitive one: a well-orchestrated team of models beats any single one. The flagship Fugu Ultra reportedly stands shoulder to shoulder with Anthropic's Fable 5 and Mythos on the hardest engineering and reasoning benchmarks. The line Sakana wants you to remember is 'frontier capability without the risk of export controls,' a not-subtle pitch to anyone nervous about being cut off from US frontier models.

This isn't a hack thrown together. It comes straight out of Sakana's TRINITY and Conductor papers at ICLR 2026, which worked out how to train a language model to route tasks to expert agents and merge their outputs. Fugu is that research turned into a product you can pay for: $5 per million input tokens, $30 per million output, with subscription tiers at $20, $100, and $200 a month.

Here's the tension worth sitting with. Fugu claims you can reach the frontier by orchestrating models the same week PlanBench-XL showed that even GPT-5.4's tool-use planning collapses from 52% to 11% the moment tools start failing. Orchestration is powerful and orchestration is fragile, both at once. Sakana is betting the powerful half wins. Release at sakana.ai/fugu-release.