Sakana upgrades 'Model Squad': Ultra v2 for performance, Max for cost reduction.

动察 Beating AI News Flash: Japanese AI company Sakana AI releases Fugu Max and Fugu Ultra v2.

Both are essentially "model orchestrators": they first assess the task, then select suitable models to divide the work, cross-check each other, and finally synthesize an answer.

The first-generation Fugu relied on this approach to team up models like GPT, Claude, and Gemini, achieving performance close to or even partially surpassing the then-stronger Fable 5.

This time the product line splits into two tracks. Fugu Max focuses on saving money, adding more open-weight and specialized models to its model pool and prioritizing models that are sufficient but cheaper. The company claims that at near-top-tier model performance, the cost is about 1/2 to 1/6 of the competitor's. Fugu Ultra v2, by contrast, pursues only maximum performance, ranking first or tied for first in 5 out of 8 benchmarks, and its Agent pool does not include Fable 5, Fable 5.1, or GPT-6 Astra.

However, switching back and forth among multiple models may reduce context cache reuse and also generate extra orchestration tokens. Sakana did not disclose Max's cache hit rate or total token cost per task. The first-generation Fugu was also criticized by users for being slow and burning through quotas quickly.