Sakana AI introduces Fugu max and Fugu ultra v2 for enhanced multi-agent orchestration

Sakana AI has launched Fugu Max and Fugu Ultra v2, two new models in its Fugu family, designed for efficient multi-agent orchestration. These models operate as orchestrators that route tasks across a pool of other models via a single API. Fugu Max focuses on cost-effectiveness, while Fugu Ultra v2 emphasizes high capability for complex tasks. Both models are available as hosted APIs through Sakana’s OpenAI-compatible service, with no open weights for self-hosting and no EU/EEA availability. The release follows a rapid development cycle, with Fugu entering beta in April 2026 and reaching general availability in June. The models are built on two ICLR 2026 papers, incorporating techniques like reinforcement learning and evolutionary algorithms. Fugu Max expands the model pool by integrating NVIDIA’s Nemotron family, routing tasks to the most efficient model. Sakana claims Fugu Max achieves elite performance at 2x to 6x lower cost, while Fugu Ultra v2 targets complex reasoning and full-stack development, with a training cutoff of August 28, 2026.