Fugu Max and Fugu Ultra v2: Orchestrating the Pareto Frontier

Rather than building a single larger model, Sakana AI is betting on orchestration: routing each incoming task to whichever model in a large pool, open or specialized, is best suited to handle it. Fugu Max expands that pool to include NVIDIA Nemotron among other open and specialized models, and Sakana claims the resulting system reaches frontier-level results at two to six times lower cost than calling a single frontier model directly for every request. A companion release, Fugu Ultra v2, aims at the opposite end of the tradeoff, pushing peak orchestrated capability as high as it goes without relying on a frontier model in the loop at all. The pairing gives teams evaluating multi-model routing against single frontier APIs a concrete cost and capability data point rather than a general argument for orchestration. The open question the release does not answer is routing overhead. Sending a task through a classifier before it reaches a model adds latency that a single well-chosen model call does not have, and that cost is not part of the headline savings figure.