Sakana Fugu Ultra is built on a different premise from typical foundation models: instead of training a single large base model, it acts as a multi-agent orchestrator that dynamically routes work across a pool of existing frontier models and exposes them through one OpenAI-compatible API. Sakana describes the approach as "One Model to Command Them All," with the Ultra variant positioned for complex, multi-step tasks that benefit from choosing and sequencing specialized models rather than relying on any one vendor. The system is distributed through several gateways, including OpenRouter, Vercel AI Gateway, and opencode, so applications can integrate orchestration without rewriting around different provider SDKs.

In practice, Fugu Ultra is aimed at engineering, scientific, and reasoning workloads where picking the right underlying model matters more than raw latency. Sakana-reported numbers cited by third-party coverage place Fugu Ultra at 73.7% on SWE-Bench Pro, ahead of Claude Opus 4.8 at 69.2% and other frontier entries, though those figures have not yet been independently reproduced. Because the value comes from coordination rather than a fixed parameter count, Fugu Ultra suits teams that want frontier-quality answers on multi-step problems while keeping a single integration point, with the caveat that the service is not yet available in the EU/EEA while Sakana works toward GDPR compliance.