News · 2026-09-13

Sakana's Fugu Max sells a model that hands your request to other models, at $2 per million tokens in

Sakana AI, the Tokyo-based research lab, launched Fugu Max on 11 September 2026: a model that does not answer most requests itself but sends each task to a pool of other models and assembles the result, sold as a single API at $2 per million input tokens and $6 per million output tokens. Sakana says it gets close to top models at two to six times lower cost, but it has not disclosed which models are in the pool.

Key facts

• Price: $2 per million input tokens and $6 per million output tokens, “Fixed rates regardless of context length,” with a 1 million token context window.
• When: released 11 September 2026, alongside a higher-end Fugu Ultra v2 at $5 and $30.
• Who: Sakana AI, with NVIDIA's Nemotron models in the pool through a collaboration with NVIDIA.
• Primary sources: the Fugu Max announcement, the product and pricing page and the OpenRouter listing.

A conductor, not a soloist

Most AI APIs give you one model. Fugu gives you a manager. As Sakana's original Fugu announcement put it, “Sakana Fugu is itself a language model trained to call various LLMs in an agent pool, including instances of itself recursively.” When Ground Truth first covered the idea in June, we described it as a committee of models behind one door.

The analogy is a general contractor. You hire one firm to renovate a kitchen, and it decides which jobs go to its own staff and which go to the electrician, plumber and tiler, then hands you a finished room and one invoice. Fugu decides which model handles which part of a request, and you pay one price.

This is a productised version of a well-studied idea, explained in our lesson on model routing and cascades: spend expensive compute only where it changes the answer.

What is new in Max

The June version of Fugu leaned on a mix of models. Fugu Max, Sakana says, “expands the pool of models Sakana Fugu can orchestrate, integrating an unprecedented number of open-weights and specialized models, including NVIDIA Nemotron family through our collaboration with NVIDIA.” Building on cheaper open-weight models is what makes the low price possible.

The pitch is explicitly about cost. “At $2 per million input tokens and $6 per million output tokens, Fugu Max's output pricing is 40-60% lower than Sonnet 5, GPT 5.6 Terra, and Kimi K3,” the announcement says, promising “performance within striking distance of elite models at two to six times lower cost.” Sakana reports the “best overall score on six benchmarks,” covering tasks such as command-line agent work, graduate-level science questions, long-document reasoning, reading PDFs, workflow automation and an internal coding test built from Sakana's own programming challenges.

Fugu Ultra v2, the premium tier, states which models it does not use: “Fable 5, Fable 5.1 and GPT-6-Astra are NOT in Fugu-Ultra v2's model pool.”

Where it runs

Fugu Max is available through Sakana's API, on OpenRouter, and on Vercel's AI Gateway, where the same $2 and $6 pricing is listed. Sakana also offers subscription plans. The company's product page notes it is “Not yet available in the EU/EEA while we work toward compliance with GDPR and EU-specific regulations.” The technical approach is described in the Sakana Fugu Technical Report, which covers the June system rather than Max specifically.

Why it matters

For developers, the question is whether a routed bundle of cheaper models can replace a frontier API for everyday agent and coding work. If it can, the value shifts from owning the single best model to orchestrating many good ones, which is a business any lab with a strong router can enter. It also lowers dependence on any single provider, which was Sakana's original pitch.

The caveat

Every score is Sakana's own, one of the headline tests is internal, and no independent evaluation has been published. Because the pool is undisclosed, customers cannot tell which models, and which licences, are handling their data, and the answers may change as Sakana swaps models in and out. Anyone considering a switch should test it on their own workload first.

Primary source, verified: read the paper →

Ground Truth. (2026, September 13). Sakana's Fugu Max sells a model that hands your request to other models, at $2 per million tokens in. Ground Truth. https://groundtruth.day/news/sakana-fugu-max-sells-a-model-that-hands-your-request-to-other-models-for-2-dollars.html

@misc{groundtruth:sakana-fugu-max-sells-a-model-that-hands-your-request-to-other-models-for-2-dollars,
  title  = {Sakana's Fugu Max sells a model that hands your request to other models, at $2 per million tokens in},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {sep},
  url    = {https://groundtruth.day/news/sakana-fugu-max-sells-a-model-that-hands-your-request-to-other-models-for-2-dollars.html}
}

Topics: sakana · model-routing · api · orchestration · pricing · open-weights