Sakana AI's Fugu Ultra v2 and Fugu Max answer by routing work to other models
Sakana AI released Fugu Max and Fugu Ultra v2 on 11 September 2026[1]. Both are served through an OpenAI-compatible API, and every Fugu model has a 1,000,000 token context[1,4].
Fugu is a multi-agent system delivered as one model[3]. For each task it picks models from a pool and switches between them[3]. Which models it uses and how it coordinates them are proprietary and not exposed[3].
Max and Ultra v2 are the same orchestration design tuned for two different aims[1]. Max sends each task to the leanest model that can solve it, to spend fewer tokens[1].
Sakana says Ultra v2 reaches its scores without Fable 5, Fable 5.1 or GPT-6 Astra in its pool[1]. Through a collaboration with NVIDIA, it does use the Nemotron family[1].
Per million tokens, fugu-ultra-v2.0 is $5 input, $0.50 cached input and $30 output[2]. Above 272K tokens of context it becomes $10, $1.00 and $45[2].
Both the 272K line and the shape of the step, with input and cache doubling while output rises by half, match gpt-6-astra[2,6]. Sakana's price list does not say whether the higher rate covers the whole request[2].
fugu-max-v1.0 is $2 input, $0.25 cached input and $6 output, regardless of context length[2]. Calls to web_search and web_fetch are $0.007 each[2].
In the announcement, Ultra v2 scores 48.3 on Chartography, against 27.3 for Opus 5 and 29.5 for Fable 5, and 74.3 on DeepSWE[1]. These are figures Sakana published itself[1].
The service is not yet available in the EU or EEA, and use of your data for training can be turned off from the console at any time[3].
The same price list carries Sakana Namazu, a model specialised for Japanese, at $0.95 input and $4.00 output[2,5].