Sakana: Fugu Max

Sakana Fugu Max is the cost-performance tier of Sakana AI's Fugu family, released on September 11, 2026. It is not a single monolithic foundation model but a learned multi-agent orchestration system: orchestrator language models read user queries, decompose work, route subtasks across a fixed pool of open-weight and specialized worker models (including NVIDIA Nemotron through Sakana's partnership with NVIDIA), and synthesize final answers. Fugu Max extends that pool to the largest configuration Sakana has shipped to date and optimizes routing for the leanest combination of agents that can still solve each task, targeting frontier-grade outcomes at roughly $2 per million input tokens and $6 per million output tokens.

The product is exposed as a standard OpenAI-compatible API (fugu-max / fugu-max-v1.0) with a one-million-token context window, configurable reasoning effort, function calling, structured outputs, built-in web search and fetch, and multimodal inputs including images and PDFs. Sakana positions Fugu Max against similarly priced single-model APIs and reports that it leads its comparison set on six of ten published benchmarks in that tier—including Terminal-Bench 2.1, GPQA Diamond, GDP.pdf, and AutomationBench—while expanding the cost-performance Pareto frontier on seven of those ten tasks.

Architecturally, Fugu builds on Sakana's Conductor and TRINITY orchestration research: a larger coordinator trained with reinforcement learning can author multi-step playbooks and recursive reasoning scaffolds, while lighter routing assigns roles across the worker pool. Because weights are not published and inference runs only through Sakana's hosted API, teams adopt Fugu Max as a vendor-managed orchestration layer rather than a self-hosted checkpoint.