Sakana Fugu (Fugu / Fugu Ultra)
A model from the Japanese lab Sakana AI, commercially launched in June 2026, that is not really a model: it is a multi-agent system disguised as one, exposed through an OpenAI-compatible API. You send your request to a single endpoint; behind it, Fugu decides on its own whether to answer itself or assemble a team of specialized frontier models, coordinate them, verify and synthesize the result. Two variants: Fugu (balanced, low latency) and Fugu Ultra (maximum quality on hard problems, mobilizing a deeper agent pool). Grounded in two ICLR 2026 papers (TRINITY and Conductor) on learned orchestration.
Strengths
- Anti-lock-in: routes around a provider that goes down or tightens its terms
- OpenAI-compatible API: wire it in with a plain URL swap, no rewrite
- Automatic routing to the specialized model best fit for the task
Limitations
- Higher latency: coordinating several models costs time versus a direct call
- Opacity: you do not choose which sub-model handles your request or sees your data
- Cost and relevance to validate: for many cases, a direct frontier model is enough
Best for
- A team whose constraint is not the score but multi-vendor service continuity
- A tech lead who wants an anti-dependency abstraction layer without rewriting their calls