Sakana AI Brings Fugu Max Routing to OpenRouter

Sakana AI has made Fugu Max available on OpenRouter under the identifier sakana/fugu-max. Rather than operating as a single model, Fugu Max is a learned multi-agent orchestration engine. It dynamically routes user requests across a pool of open-weights and specialized models, such as the NVIDIA Nemotron family, and can even recursively call itself to break complex tasks into smaller subtasks. The system features a one-million-token context window, supports image and PDF inputs, and offers web search, function calling, and structured outputs.

The service is priced at $2 per million input tokens and $6 per million output tokens, which Sakana AI claims is 40% to 60% cheaper than frontier models like Sonnet 5, GPT 5.6 Terra, and Kimi K3. Alongside Max, Sakana released Fugu Ultra v2 for complex, multi-step agent workflows. Ultra v2 achieved a score of 48.3 on Chartography and 74.3 on DeepSWE without relying on closed frontier models like Fable 5, Fable 5.1, or GPT-6-Astra. Meanwhile, Fugu Max recorded top marks across six evaluations, including Terminal Bench 2.1, GPQAD, AA-LCR, GDP.pdf, AutomationBench, and SWEFish, while beating Opus 5 at 27.3 and Fable 5 at 29.5 on specific tests.

For AI developers, Fugu Max replaces manual, application-level routing rules with an automated, execution-time decision process. However, this orchestration overhead introduces distinct trade-offs. OpenRouter telemetry shows a median P50 latency of approximately 5.4 seconds and a throughput of around 13 tokens per second. These speeds make the engine less suitable for real-time interactive chat or autocomplete, but highly viable for background agent loops, terminal tasks, and long-document analysis. Practitioners must also account for the fact that total billing includes the tokens consumed by the routing and synthesis processes themselves.