Sakana AI Launches Fugu Max and Fugu Ultra v2, Models That Orchestrate Other Models

The Tokyo-based laboratory relies on a learned orchestrator rather than a single model: Fugu routes tasks to a pool of models and claims better scores for less.

On September 10, 2026, Sakana AI posted two new members of her Fugu family: Fugu Max, cost oriented, and Fugu Ultra v2, quality-oriented. The principle departs from the race to the monolithic model: Fugu is a trained orchestrator who divides the work between a pool of models behind a single API, and can be called recursively. The approach is based on two works presented to ICLR 2026, TRINITY, who assigns roles as thinker, executor or auditor, and The Conductor, trained by strengthening to coordinate these roles in natural language.

Fugu Max is billed $2 per million tokens in admission and $6 in release, 40 to 60% less than Sonnet 5, GPT 5.6 Terra or Kimi K3, while claiming the best overall scores on Terminal Bench 2.1, GPQA Diamond, AA-LCR, GDP.pdf, AutomationBench and SWEFish. Fugu Ultra v2 aims for long reasoning, stand-alone research and complete software development tasks: 48.3 on Chartography, a test of visual reasoning and data interpretation, compared to 27.3 for Opus 5, and 74.3 on DeepSWE against three to five times more expensive models at the token.

A detail is worth noting: the Fugu Ultra v2 model pool excludes Fable 5, Fable 5.1 and GPT-6 Astra, making comparisons less obvious to read than it appears. Both models are immediately accessible via an OpenAI compatible API, without open weights, and are not available in the European Union or in the European Economic Area.