Sakana AI has released Fugu Max and Fugu Ultra v2 – two new models from the Sakana Fugu family. Fugu is not just a single model, but a trained orchestrator that distributes tasks among many other models through a single API. Fugu Max is aimed at the best output per dollar, while Fugu Ultra v2 targets the highest performance in solving complex multi-stage tasks. Both models are available through Sakana's OpenAI-compatible API. There are no open weights for self-deployment, and Sakana does not offer the service in the EU/EEA.
Sakana considers the task as two-dimensional: real workloads are evaluated by capability and cost together. On the Pareto frontier, improving quality is more expensive, and reducing price leads to a loss of quality. Fugu Max and Fugu Ultra v2 have the same orchestration architecture – only the optimization goals differ. The release is rapid: Fugu entered beta in April, reached general availability in June, and in July received Fugu-Cyber and the Claude Code interface.
According to the technical report, Fugu models are essentially language models: they read the request and create an agent construction for it on the fly. Training combines large-scale fine-tuning, evolutionary algorithms, and reinforcement learning. The system relies on two ICLR 2026 papers: TRINITY uses a lightweight evolved coordinator that assigns roles of Thinker, Worker, or Checker by turns, and Conductor is trained with reinforcement learning to discover natural language coordination strategies.
Fugu Max expands the set of models that Fugu can orchestrate by adding open weights and specialized models, including the NVIDIA Nemotron family thanks to a collaboration with NVIDIA. Each task is directed to the most cost-effective model capable of solving it. Prices: $2 per 1 million input tokens and $6 per 1 million output tokens – 40–60% lower than Sonnet 5, GPT 5.6 Terra, and Kimi K3. Fugu Max demonstrates the best overall result on six benchmarks (Terminal Bench 2.1, GPQA Diamond, AA-LCR, GDP.pdf, AutomationBench, and SWEFish) and extends the cost-output frontier on 7 out of 10. SWEFish is Sakana's internal benchmark created on proprietary programming tasks, so its result should be considered a vendor signal.
Related articles
