Sakana AI launches Fugu — one model to command them all

Sakana AI — the Tokyo lab founded by David Ha and Llion Jones — has launched Sakana Fugu, a language model that doesn't just answer prompts but orchestrates other models. Announced on June 22, 2026, Fugu is itself an LLM trained to coordinate a swappable pool of frontier models: it decides whether to answer directly or delegate, how to split work across "Thinker / Worker / Verifier" roles, and how to aggregate and verify the results. It can even invoke itself recursively.

#The pitch: resilience over lock-in

Sakana frames Fugu as an answer to vendor and geopolitical risk. Because the underlying agent pool is interchangeable, Fugu can route around a provider that becomes unavailable.

"As we have seen recently from export controls imposed on Anthropic's Fable and Mythos models, access can shift or disappear overnight due to changing regulatory boundaries, export controls, and foreign policies."

"For an organization or a nation, relying on a single company's APIs for critical infrastructure, finance, or governance is a material vulnerability."

#Two variants, one API

• Fugu — balances quality and latency; the default for everyday coding, code review and chat.
• Fugu Ultra — tuned for maximum accuracy on demanding multi-step problems, coordinating deeper expert pools.

Both sit behind a single OpenAI-compatible API (console.sakana.ai). The work builds on two ICLR 2026 papers from the lab — Trinity (a lightweight evolved coordinator) and The Conductor (an RL-trained system that discovers natural-language coordination strategies).

#Benchmarks (as reported)

Sakana says Fugu Ultra "stands shoulder-to-shoulder" with Fable 5 and Mythos Preview. Reported figures for Fugu Ultra include SWE-Bench Pro 73.7%, TerminalBench 2.1 82.1%, LiveCodeBench 93.2%, Humanity's Last Exam 50.0% and CharXiv Reasoning 86.6%. These come largely from secondary coverage of the launch charts, so treat the exact digits as approximate.

Fugu is proprietary and API-only (not open weights), offered via subscription tiers plus pay-as-you-go pricing for heavier workloads.

Read the announcement: Sakana Fugu: One Model to Command Them All.