Sakana AI's Fugu orchestrates multiple AI models via a single API

Sakana AI, the Tokyo research lab known for evolutionary model-merging and the 'AI Scientist' work, has moved its multi-agent orchestration system Fugu from beta to commercial launch. Fugu presents itself as a single OpenAI-compatible endpoint that internally selects, delegates, verifies, and synthesizes answers from a pool of frontier models — including itself. This eliminates the need for teams to manually wire up multi-agent architectures using frameworks like LangGraph or CrewAI.

Two tiers are available: Fugu for balanced performance and latency, and Fugu Ultra for hard, high-stakes problems like Kaggle competitions, paper reproduction, cybersecurity, and patent search. Both use the same API, allowing seamless switching. Pricing is $20/$100/$200 per month for subscriptions (both models included), or pay-as-you-go for Fugu Ultra at $5 input / $30 output per 1M tokens. Fugu is not yet available in the EU/EEA while Sakana works toward GDPR compliance.

Sakana positions Fugu as 'shoulder-to-shoulder' with limited-access frontier systems like Anthropic's Fable 5 and Mythos Preview, while delivering frontier capability without export control risks — a pitch wrapped in geopolitical argument. However, benchmarks are vendor-reported and not a clean sweep: Fugu Ultra leads several reasoning and coding benchmarks, but Fable 5 tops SWE-Bench Pro, GPT-5.5 leads MRCRv2, and Opus 4.8 leads the CTI-REALM security benchmark. The headline claim of 'AI sovereignty' — routing around provider restrictions — is real but still relies on renting intelligence from those same vendors.