Sakana AI Releases Fugu Max and Fugu Ultra v2
Sakana AI has released Fugu Max and Fugu Ultra v2, two models in its Fugu family. Fugu is a learned orchestrator rather than a single foundation model: it reads a request and routes work among other models through one API. Both releases use the same orchestration architecture, but Fugu Max is optimized for cost efficiency while Fugu Ultra v2 targets difficult, multi-step tasks.
Fugu Max can select from a broader pool of open-weight and specialized systems, including NVIDIA Nemotron models. Sakana AI lists pricing of $2 per 1M input tokens and $6 per 1M output tokens. The company reports the best overall results on six benchmarks and says the model expands the cost-performance Pareto frontier on seven of ten benchmarks. SWEFish is an internal benchmark based on Sakana’s coding challenges.
Fugu Ultra v2 is aimed at complex reasoning, autonomous research, and full-stack software development. Sakana AI reports scores of 48.3 on Chartography and 74.3 on DeepSWE, plus the best or joint-best result on five of eight benchmarks. Neither model has open weights, and the hosted, OpenAI-compatible API is unavailable in the EU/EEA. Fugu Ultra v2’s training cutoff is August 28, 2026.
Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.