Sakana AI launches Fugu, an orchestration model that commands a pool of LLMs
On 22 June 2026, Tokyo-based Sakana AI released Sakana Fugu, a product that packages a full multi-agent orchestration system inside a single foundation model. Rather than chasing scale with one ever-larger monolith, Fugu is itself a language model trained to call other LLMs from a swappable agent pool — including instances of itself recursively — planning, delegating, verifying and synthesising behind one OpenAI-compatible API.
It ships in two tiers. Fugu balances performance with low latency as an everyday default, slotting into tools such as Codex for coding and code review, and lets teams opt specific agents out of the pool for privacy or compliance. Fugu Ultra is tuned for maximum answer quality on hard, multi-step problems, coordinating a deeper pool of expert agents; early users have applied it to AI research, paper reproduction and cybersecurity analysis. Sakana claims Fugu Ultra stands shoulder-to-shoulder with frontier models such as Anthropic's Fable 5 and Mythos Preview across engineering, scientific and reasoning benchmarks.
The pitch extends beyond capability to resilience. Sakana frames orchestration as a hedge against single-vendor dependency and shifting export controls — citing recent restrictions on Anthropic's Fable and Mythos models — arguing that if one provider's access disappears, Fugu dynamically routes around the disruption. The work builds on the company's research on learned model orchestration, including its ICLR 2026 papers Trinity and Conductor.
Read more: Sakana AI · Sakana AI on X
