← Back to Blog
New model

Sakana AI Fugu Max and Fugu Ultra v2 API: orchestration models

FastMetal

Fugu Max and Fugu Ultra v2, released by Tokyo's Sakana AI on September 11, are on FastMetal as fugu-max and fugu-ultra-v2.

Fugu is an orchestrator. According to Sakana AI's announcement, it has learned which model to use for which task, hands the work to several models behind it, and assembles their output into one answer. From the caller's side it is a normal Chat Completions API.

Max or Ultra v2

Fugu Max is the cost-focused tier and Fugu Ultra v2 the performance tier. Both have:

  • A 1,000,000-token context and up to 128,000 output tokens.
  • Image input, tool calling with tools, and structured output with a strict json_schema. We confirmed all three with real calls through FastMetal.

Sakana AI also sells a web search feature billed per use. FastMetal does not accept per-use add-ons of that kind, so web search is unavailable through us.

Pricing

Both bill per million input tokens and per million output tokens, with a lower rate for cached input. Ultra v2 costs more than Max, and the gap is widest on output. Current rates are on the Fugu Max and Fugu Ultra v2 model pages.

A sensible pattern: start with Max, and switch to Ultra v2 only where Max falls short.

Ultra v2 bills about 2,000 input tokens for a one-line question

This showed up in our first calls and it affects the bill directly. Ultra v2 adds the orchestrator's own prompt to whatever you send. A one-line question ("What is the capital of Japan? One word.") billed 1,944 input tokens.

That fixed part caches. Sent again right away, 1,942 of the 1,944 input tokens counted as cache reads, and the cost of the request dropped sharply. For a stream of short requests, every call after the first gets that benefit.

Fugu Max billed 87 input tokens for the same question, so its fixed part is barely visible. It does think, though: 132 of its 133 output tokens were reasoning. A one-word answer still bills for the reasoning behind it.

Calling it

Pass fugu-max or fugu-ultra-v2 as the model:

curl https://api.fastmetal.ai/v1/chat/completions \
  -H "Authorization: Bearer $FASTMETAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "fugu-max",
    "messages": [{"role": "user", "content": "Explain the cause of this error log in three lines."}]
  }'

Send images as base64 data URLs. In our tests, a request with an image URL on an outside site returned a 400 on both models, and the same image sent as base64 was read correctly.

curl https://api.fastmetal.ai/v1/chat/completions \
  -H "Authorization: Bearer $FASTMETAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "fugu-ultra-v2",
    "messages": [{
      "role": "user",
      "content": [
        {"type": "text", "text": "Summarize this diagram."},
        {"type": "image_url", "image_url": {"url": "data:image/png;base64,iVBORw0KGgo..."}}
      ]
    }]
  }'

To see both models answer the same prompts, open the Fugu Max vs Fugu Ultra v2 comparison. For other models, see the model catalog and pricing.

Try Fugu Max right now

Fugu Max is available on FastMetal through one API key. Start in the browser, or call it from the OpenAI SDK.