Sakana AI released two models — Fugu Max and Fugu Ultra v2. According to MarkTechPost, they are built on the same orchestration architecture trained on examples. Fugu Max directs tasks to compact open and specialized models, including NVIDIA Nemotron, at a price of 2 and 6 dollars for 1 million tokens.
Fugu Ultra v2 is oriented toward maximum capabilities. In the material provided, its results are 48,3 in Chartography and 74,3 in DeepSWE.
The news matters in that Sakana AI simultaneously claims economical and high-performance scenarios for multi-agent orchestration. At the same time, the source is presented only with metadata and a synopsis of the publication, so independent confirmation of the claimed characteristics is absent.
editorial commentary
Why it matters
A likely implication is that users will be offered a choice between cheaper routing and maximum claimed performance. The nearest verifiable signal is Sakana AI's primary materials with descriptions of pricing, tests and launch conditions. A significant uncertainty is the lack of primary confirmation and details of the methodology.