Introducing Fugu Max and Fugu Ultra v2: Orchestrating the Pareto Frontier
Sakana AI ● Covered by 2 sources
Sakana AI launched Fugu Max and Fugu Ultra v2. One chases lower cost; the other pushes higher performance without leaning on the usual frontier models.
Based on reporting by Sakana AI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Sakana AI is making a simple argument: the race in AI should not be about bigger models alone. It should be about getting the right model for the right job at the lowest cost, then using orchestration to stitch those choices together.
That’s the pitch behind Fugu Max and Fugu Ultra v2. They share the same orchestration architecture, but they are tuned for different goals. Fugu Max broadens the pool of models Fugu can route work to, adding an unusually large set of open-weights and specialized models, including NVIDIA’s Nemotron family through a collaboration with NVIDIA.
The company says that setup lets Fugu Max send tasks to the leanest model that can still handle them, producing frontier-level results while cutting token spend. Sakana also says the system reaches a point on the Pareto frontier that single-model providers cannot match, with performance close to elite models at two to six times lower cost.
Fugu Ultra v2 is the other side of the same coin. It is aimed at harder work: multi-step reasoning, autonomous research, and full-stack software development. Sakana says it performs especially well on complex visual and structured data, pointing to a 48.3 score on Chartography, ahead of Opus 5 at 27.3 and Fable 5 at 29.5, and a 74.3 score on DeepSWE, ahead of models that cost three to five times more per token.
The company also says Ultra v2 gets those results without Fable 5, Fable 5.1, or GPT-6-Astra in its agent pool. Both Fugu Max and Fugu Ultra v2 are available now through an OpenAI-compatible API, and users already on Fugu only need a single-line parameter change to upgrade. Sakana’s bigger point is clear: the winning system may not be one giant model at all, but a swarm of smaller ones that can be swapped in and out without drama.
My take — AI-written commentary, not fact-checked reporting
This is the sane direction for AI, which is exactly why it keeps getting dressed up as a revolution. Single-model worship is a nice way to sell expensive tokens; orchestration is how grown-up systems get built. The open-model angle matters too, because vendor lock-in is just a fancy term for handing your plumbing to someone else and hoping they stay polite.
Read more about this at: Sakana AI
