Sakana Fugu Ultra

Fugu Ultra launched alongside Fugu on 22 June 2026 as the quality-first variant of Japan's Sakana AI orchestration engine: where Fugu optimises for latency, Ultra coordinates a deeper pool of expert agents when accuracy and depth matter more than speed. Sakana's own framing of the value is instructive — not a better answer to one prompt, but sustained progress across many steps: reading, implementing, testing, comparing evidence, finding gaps and producing a final analysis. Beta users, close to 500 of them, applied it to paper reproduction, cybersecurity assessment, code review, and literature and patent investigation; one ran it in a near-fully automated research mode. The vendor's marketing claim is that Ultra stands alongside Anthropic's Fable 5 and Mythos Preview on engineering, scientific and reasoning benchmarks while carrying no export-control risk — and notes that neither of those two models is in Fugu's agent pool, because they are not publicly accessible. That caveat matters when reading the comparison: Sakana published its benchmark results as image grids rather than as figures in the text, took competitors' scores from their own vendors, and used the higher of the two Anthropic results where both were available. A technical report (arXiv 2606.21228) and two ICLR 2026 papers, Trinity and Conductor, document the learned-orchestration research behind it. No parameter count, context window or per-token price was published for this launch pair. Fugu Ultra v2, from September 2026, is a separate catalogue entry.

▸Videos

No videos yet.

▸Reviews