Japanese AI startup Sakana AI has officially launched its multi-agent system Sakana Fugu, ending beta and offering an OpenAI-compatible API to the public. The system comes in two versions: a low-latency Sakana Fugu Mini and a heavy-reasoning Sakana Fugu Ultra.
In the face of recent US export controls on models like Fable and Mythos, Sakana AI argues that reliance on a single model has become a serious security vulnerability. Fugu doesn't depend on any one foundation model — instead, it routes requests through a pool of swappable "worker" models, bypassing export restrictions and giving organizations and nations AI sovereignty through dynamic orchestration. Since Fable 5 and Mythos Preview aren't publicly available, they aren't in Fugu's worker pool, but they serve as benchmark targets.
The system architecture is based on the TRINITY and Conductor papers the team published at ICLR 2026. While the commercial version doesn't disclose base model parameters, the papers reveal an academic version powered by a 7B parameter controller model trained with reinforcement learning. The system supports adaptive recursion: models can check their own outputs and trigger corrections, and users control compute costs by adjusting recursion depth.
Benchmark results show Fugu Ultra trading blows with cutting-edge single models. On LiveCodeBench and Terminal Bench 2.1, Fugu Ultra scored 93.2 and 82.1, beating the unreleased Fable 5's 89.8 and 80.4. However, on SWEBench Pro, Fugu Ultra hit 73.7, falling short of Fable 5's 80.0. In most tests, Fugu Ultra also outperformed flagship models like GPT 5.5, Gemini 3.1 Pro, and Opus 4.8.