Sakana AI Releases Fugu Max and Ultra v2 Models

Sakana AI has released two new models, Fugu Max and Fugu Ultra v2, which are enhancements to its Fugu family of orchestrators. Fugu is designed not as a single foundation model, but as an intelligent system that routes tasks to an array of other models, accessible through a single API. The latest versions are specifically tuned for distinct objectives: Fugu Max prioritizes output quality relative to cost, aiming for the best performance per dollar spent, while Fugu Ultra v2 is engineered for maximum capability in executing difficult, multi-step tasks. Both models are available immediately via Sakana's OpenAI-compatible API. Sakana AI does not offer open weights for self-hosting and has stated that the service is not available in the European Union or European Economic Area. The company frames the development of these models as addressing a "two-axis problem" where real-world applications are evaluated on both their capability and their associated costs. Sakana's rationale is that assigning a highly complex, multi-trillion-parameter model to a simple data lookup task is an inefficient use of resources. A more effective system, according to Sakana, selects the most economical model capable of successfully completing the task. This concept is illustrated using the Pareto frontier, a graphical representation where improvements in quality typically incur higher costs, and cost reductions lead to a decrease in quality. Fugu Max and Fugu Ultra v2 share a common core orchestration architecture, with their differentiation stemming solely from their respective optimization targets. This release follows a rapid development cycle for Sakana AI; the Fugu orchestrator entered beta in April, achieved general availability in June, and subsequently introduced Fugu-Cyber and a Claude Code interface in July. The technical report from Sakana AI describes Fugu models as sophisticated language models in their own right. Upon receiving a query, they dynamically construct an "agentic scaffold" for it. The training methodology for these models integrates large-scale fine-tuning, evolutionary algorithms, and reinforcement learning. The underlying system is built upon research presented in two papers from ICLR 2026: TRINITY, which employs a lightweight, evolved coordinator to assign roles such as Thinker, Worker, or Verifier across conversational turns, and The Conductor, trained using reinforcement learning to discover effective natural-language coordination strategies and focused prompts. Fugu Max specifically expands the range of models that Fugu can orchestrate by incorporating a broad selection of open-weight and specialized models, aiming to provide greater flexibility and cost-effectiveness for a wider array of tasks.