Fugu Max and Fugu Ultra v2 bring a replaceable pool of models behind one interface. Max aims to select the leanest model able to solve a task, while Ultra v2 targets demanding work. Sakana AI is offering both through its existing Fugu interface; the reported performance results are the company's own measurements.
AI company Sakana AI has released Fugu Max and Fugu Ultra v2. Both use Fugu, an architecture that combines replaceable open and specialized models within one service. Max routes each task to the leanest model capable of solving it. Ultra v2 focuses on demanding work. Users reach the coordinated pool through a single interface.[1], [2]
The company reports that Max leads the overall scores on six benchmarks, including SWEFish, its internal test of coding tasks. For Ultra v2, it reports 48.3 on Chartography, which assesses visual reasoning, and 74.3 on the software-task benchmark DeepSWE. These are Sakana AI's measurements; no independent reproduction accompanies them. The two options target different priorities within the same architecture.[1]
The new options use Fugu's existing OpenAI-compatible application interface. Sakana AI says an existing deployment can switch with a one-line parameter change. The underlying pool is replaceable rather than fixed to one model. The company also states that Fable 5, Fable 5.1 and GPT-6-Astra are absent from the Fugu Ultra v2 pool used for its reported results.[1]
