Sakana AI released Fugu Max, the cost-performance member of its Fugu model family. It is listed on OpenRouter and hosted by a single provider. Fugu is not one model. It is a language model trained to route tasks across a fixed pool of open-weight and specialist models, including NVIDIA's Nemotron family. It can also call copies of itself. Sakana says the system picks efficient combinations of expert models to improve quality and cost together. Pricing is flat regardless of how long the prompt is. Tokens used by the routing system itself are billed as ordinary input and output tokens. The model takes text, images and PDFs, and returns text. It supports function calling, structured outputs, web search and web fetch.
What changed
Sakana's earlier text models were Fugu Ultra, Fugu Ultra v2 and Namazu.
What it unlocks
Sending million-token prompts at a flat price, with the system picking expert models per task.
- $2/M input, $6/M output
- 1M token context, 128K max output
- cache read $0.25/M
- web search $10 per 1K calls
What you need to act on it
- an OpenRouter or Sakana API account
Sources