Sakana AI this week released Fugu Ultra v2 and a new cost-optimized sibling, Fugu Max — two launches that together mark the second generation of the Japanese startup's bet on orchestration instead of scale.
A router, not another frontier model
Unlike the industry's dominant approach — training a single, ever-larger model —, Sakana describes Fugu as a learned orchestration system: a model trained to decide, task by task, which system in a pool of open-weight and specialized models should handle each step of a problem, including recursively calling instances of itself. The pool includes open models such as NVIDIA's Nemotron family and other specialized systems — but, according to the company, does not rely on any proprietary frontier model to work.
One notable detail: Sakana's published evaluation for Fugu Ultra v2 deliberately excludes models like Claude Fable 5, Claude Fable 5.1 and GPT-6 Astra from the pool of "workers" used in the test — yet still claims to outperform them on part of the benchmarks.
Reported performance and pricing
According to Sakana's own evaluation, Fugu Ultra v2 is best or joint-best on five of eight internal benchmarks, notably scoring 48.3 on "Chartography" (a visual reasoning and data-interpretation benchmark) against 27.3 for Claude Opus 5 and 29.5 for Claude Fable 5, plus 74.3 on DeepSWE. As with any benchmark released by the lab itself, it's worth waiting for independent evaluations before treating the numbers as final.
Fugu Ultra v2 targets complex multi-step reasoning, autonomous research and full-stack software development, priced at $5 per million input tokens and $30 per million output tokens. Fugu Max, optimized for cost, charges $2 input and $6 output per million tokens for teams prioritizing volume over peak performance. Rates increase past 272K tokens of context, and Sakana has not disclosed a parameter count or the orchestrator's own context-window size.
What it means for AI builders
For AI engineering teams, Fugu reinforces a trend already visible in other recent releases — such as DeepSeek V4.1-Flash — of treating orchestration and model routing as part of the architecture itself, not an external layer. Instead of picking "which model to call" when wiring up an API, developers increasingly delegate that decision to the system itself, task by task — a pattern likely to become more common as AI builds combine several specialized models rather than one generalist.