Sakana AI splits its Fugu model in two -- one built for cost, one for peak performance
Fugu Max and Fugu Ultra v2, released September 11, route each request across a pool of open-weight and specialized models rather than relying on one frontier model, undercutting Sonnet 5 and GPT-5.6 Terra on price.
Sakana AI released Fugu Max and Fugu Ultra v2 on September 11, splitting its Fugu orchestrator line into two variants that optimize for different ends of the cost-performance curve. Neither is a single trained model in the conventional sense: both are orchestration systems that take one API request and route the underlying work across a pool of open-weight and specialized models -- including Nvidia's Nemotron -- stitching the results back into one answer.
Two models, two jobs
Fugu Max is the cost-efficiency variant, priced at $2 per million input tokens and $6 per million output tokens -- 40 to 60% cheaper on output than Anthropic's Sonnet 5, GPT-5.6 Terra, and Kimi K3, according to Sakana's own comparison. The company says it posts the best score on six benchmarks including Terminal Bench 2.1 and SWEFish, and expands the cost-performance Pareto frontier on seven of ten benchmarks tested. Fugu Ultra v2, at $5 input / $30 output per million tokens, draws on a deeper expert pool aimed at complex reasoning, autonomous research, and software engineering, posting best or joint-best scores on five of eight benchmarks -- including 48.3 on the visual-reasoning benchmark Chartography and 74.3 on DeepSWE.
The pool doesn't include the current frontier leaders
The detail worth noting is what's not in Fugu Ultra v2's model pool: Sakana's own release specifically excludes Claude Fable 5, Fable 5.1, and GPT-6-Astra -- the current frontier flagships -- from the set of models Fugu routes work to. The pitch is architectural independence rather than parity with any single lab's best model: swappable components instead of a dependency on one proprietary frontier system, which Sakana frames as both a cost lever and a hedge against export-control exposure on any one vendor's chips or weights.