Fugu Ultra v2, released by Sakana AI on 11 September 2026, is the performance branch of the same orchestration architecture as Fugu Max: a pool of open-weight and specialised models, coordinated for complex multi-step reasoning, autonomous research and software work. Its distinguishing claim is a negative one — Sakana states that Fable 5, Fable 5.1 and GPT-6-Astra are not in its agent pool, so the results are reached without the strongest proprietary models on the market. Vendor figures put it at 48.3 on Chartography, a visual and structured-data reasoning test, against 27.3 for Opus 5 and 29.5 for Fable 5, and at 74.3 on DeepSWE; Sakana reports best or joint-best scores on five of eight benchmarks and a top-two place on seven of eight. None of this has been independently reproduced. Pricing is fixed per token rather than per model used, and rises above a 272K-token context, which is the only published hint at the size of the window. A technical report accompanies the release (arXiv 2606.21228).