Sakana Fugu

Fugu is the low-latency member of the pair that Japan's Sakana AI made generally available on 22 June 2026, after a beta run with close to 500 early users. It is not a conventional foundation model: Sakana describes it as a language model trained to call other language models, keeping a pool of swappable agents and deciding per request whether to answer directly or to assemble, delegate to, verify and synthesise a team of expert models — including, recursively, instances of itself. The whole multi-agent machinery sits behind a single OpenAI-compatible endpoint, so none of it reaches the caller's code. Sakana positions Fugu as the everyday default, fitted to coding tools, code review and interactive services, and it allows teams with privacy or compliance duties to opt named agents out of the pool. The vendor argues the design is also a hedge against single-vendor dependency, citing export controls imposed on Anthropic's Fable and Mythos models: if a provider cuts access, Fugu routes around it. Benchmark comparisons were published as image grids rather than figures in text, alongside a technical report (arXiv 2606.21228) and two ICLR 2026 papers, Trinity and Conductor, on learned orchestration. Sakana published no parameter count, no context window and no per-token price for this launch pair; access was sold as subscription tiers plus a pay-as-you-go plan. The later Fugu Max and Fugu Ultra v2, both from September 2026, are separate catalogue entries.

▸Videos

No videos yet.

▸Reviews