Sakana AI has released Fugu Max and Fugu Ultra v2, two new models designed for multi-agent orchestration. Both systems are currently available via a hosted API compatible with OpenAI, though they are not accessible in the EU or EEA and do not offer open weights for self-hosting.
The models utilize a shared orchestration architecture based on research from two ICLR 2026 papers, including the TRINITY framework. The training process integrates large-scale fine-tuning, reinforcement learning, and evolutionary algorithms to assign specific roles like Thinker, Worker, or Verifier to agents.
Fugu Max expands the range of models available for orchestration, including the NVIDIA Nemotron family. According to Sakana AI, Fugu Max performs near elite-level models while reducing operational costs by two to six times, based on internal benchmarks like SWEFish.
Fugu Ultra v2 focuses on complex reasoning, autonomous research, and software development, with improvements in processing visual and structured data. The model has a training cutoff date of August 28, 2026, and excludes specific models like Fable 5 and GPT-6-Astra.
Sakana AI emphasizes that its orchestration approach aims to prevent vendor lock-in and mitigate risks associated with API revocations or service interruptions. The company has maintained a rapid release schedule, having launched the Fugu beta in April 2026.
