Sakana AI releases Fugu Max and Fugu Ultra v2 orchestration models

Sakana AI has released two new versions of its orchestration platform, Fugu Max and Fugu Ultra v2, designed to manage multiple AI models to optimize for either cost or performance.

Fugu Max is a cost-efficient version that aims to provide the best possible output at the lowest price. It utilizes a dynamic routing system to assign tasks to the lightest capable model, including the NVIDIA Nemotron family. It features an input cost of $2 and an output cost of $6 per million tokens, which the company claims is 40% to 60% lower than competitors like Sonnet 5 and GPT 5.6 Terra. Fugu Max achieved top scores in six benchmarks, including Terminal Bench 2.1 and GPQAD.

Fugu Ultra v2 is optimized for high-performance, complex, multi-step tasks such as autonomous research and full-stack software development. In the Chartography benchmark for visual reasoning and data interpretation, it scored 48.3, surpassing Opus 5 (27.3) and Fable 5 (29.5). It also recorded a score of 74.3 in the DeepSWE software engineering benchmark. The platform uses an interchangeable pool of open-weight and specialized models to protect users from vendor lock-in and service disruptions.