Sakana Fugu Max Targets Frontier AI at Lower Cost
Sakana AI released Fugu Max and Fugu Ultra v2, using model orchestration to pursue cheaper agent performance and higher autonomous-task scores.
Two releases, one orchestration strategy
Sakana AI has released Fugu Max and Fugu Ultra v2, two versions of its multi-agent orchestration system aimed at different points on the capability-cost curve. Fugu Max is designed to route each task to the least expensive model capable of solving it, while Fugu Ultra v2 is optimized for maximum performance on demanding, multi-step workloads.
The system combines open-weight and specialized models rather than depending on one proprietary frontier model. Sakana says the pool now includes NVIDIA’s Nemotron family and can be swapped as models, prices, or access conditions change. Both releases are available through an OpenAI-compatible API, allowing existing Fugu users to switch with a parameter change.
Sakana reports that Fugu Max scores best overall on six internal or external benchmarks, including Terminal Bench 2.1, GPQAD, AutomationBench and SWEFish. It lists pricing of $2 per million input tokens and $6 per million output tokens, claiming performance within reach of leading models at two to six times lower cost. Fugu Ultra v2 is positioned at the high end: Sakana reports a 48.3 score on Chartography and 74.3 on DeepSWE, ahead of the comparison models cited in its release.
Those results remain vendor-reported, and the model pool excludes newer systems such as GPT-6 Astra and Claude Fable 5.1. That makes direct comparisons difficult, especially for tasks where routing overhead, tool reliability and latency matter as much as benchmark accuracy.
Why it matters
The important shift is architectural. If Sakana’s numbers hold outside its own evaluation setup, model orchestration could become a practical way to reduce dependence on a single frontier provider while preserving strong agent performance. The unresolved question is whether a dynamic collection of smaller models can remain dependable in production, where failures in routing or coordination may erase the apparent savings.