Model comparison · 2026-09-11
What Is Sakana Fugu Max? Multi-Agent Orchestration Explained
Sakana Fugu Max is not a single model but a learned orchestrator that spreads work across a pool of expert agents. Read how it works, what the published benchmark claims say, which capabilities LLMTR measured, and how it differs from Fugu Ultra.
What is Fugu Max?
Sakana Fugu Max is not a single language model. It is an orchestrator trained to take an incoming request, break the work apart, and route each part to an expert model in a pool. The agents' outputs are merged and returned to you as one OpenAI-compatible response; there is no extra protocol on the calling side.
The pool leans on open-weights models, and Sakana names the NVIDIA Nemotron family among them. Which request goes to which agent is not published: the routing decision is closed on the vendor side. Fugu Max should therefore be evaluated as a decision layer, not as a list of models.
In practice that means an easy request and a hard one sent under the same model identifier can run teams of different sizes behind the scenes. That is the upside on quality. On cost it opens a separate question, and that question is answered with measured figures in this series' pricing article.
Model card
The values below are the catalog row itself. LLMTR pins the upstream identifier to a version: a call to sakana/fugu-max reaches fugu-max-v1.0 on the Sakana side. The vendor's fugu-max shortcut is an alias and may point somewhere else later, while the price card is tied to the version. Pinning prevents a silent drift between the card we bill and the model that actually runs.
The public identifier does not change: you always call sakana/fugu-max.
| Field | Value |
|---|---|
| LLMTR model identifier | sakana/fugu-max |
| Sakana model identifier | fugu-max-v1.0 |
| Context window | 1,000,000 tokens |
| Input price | $2 / 1M tokens |
| Output price | $6 / 1M tokens |
| Cache read | $0.25 / 1M tokens |
| Endpoints | /v1/chat/completions, /v1/responses |
| Reasoning levels | high, xhigh |
Published benchmark claims and how to read them
Sakana reports that Fugu Max takes the best overall score on six benchmarks: Terminal Bench 2.1, GPQA Diamond, AA-LCR, GDP.pdf, AutomationBench and SWEFish. The same page states that it expands the cost-performance frontier on seven out of ten benchmarks.
These are the vendor's own measurements and have not been independently verified. A benchmark table is not evidence that the model will produce the same result on your work. When evaluating a new model, compare candidates on the same tasks, the same source material and the same acceptance criteria.
- Put an ordinary task next to a hard one; a model has to work in the average case, not only at the edge.
- Measure the finished job rather than the first response: an answer that needs repair costs both time and tokens.
- Record total cost including retries; a low token price alone does not mean a low cost per completed task.
Capabilities: measured, not quoted
None of the rows below were taken from a vendor feature table; each was exercised against fugu-max-v1.0 on 11 September 2026.
The json_object difference matters in particular. Fugu Max returns plain JSON. Fugu Ultra accepts the same request without an error but wraps the answer in a fenced code block, so the caller only discovers the problem at parse time. That is why the Fugu Ultra row carries no json_mode capability in the catalog, and why structured output on that model goes through json_schema.
| Capability | Status | How it was verified |
|---|---|---|
| Image input | Yes | A real PNG was sent; the returned description matched the image |
| Function calling | Yes | tool_choice: required returned a well-formed tool_calls array |
| Structured output (json_schema) | Yes | A strict schema was requested and parsed as exactly that object |
| JSON mode (json_object) | Yes | Bare JSON came back, with no fenced code block around it |
| Prompt cache | Yes | A repeated prefix reported cached_tokens above zero |
| Reasoning effort | high, xhigh | low and medium are rejected with 400; max is an alias for xhigh |
| Published output ceiling | None | The max_tokens field is not validated; no ceiling is advertised |
Fugu Max or Fugu Ultra?
Both rows use the same agent approach; they diverge on price, context banding and structured-output behaviour. Fugu Max is the default for most coding, research and analysis work. Where answer quality outranks cost, particularly on long multi-step problems, Fugu Ultra is the option.
The price gap is not small: the same request costs 2.5 times more on input and 5 times more on output on Fugu Ultra. Fugu Ultra also moves into a second price band above 272,000 input tokens, whereas the Fugu Max rate is independent of context length.
| Comparison | sakana/fugu-max | sakana/fugu-ultra |
|---|---|---|
| Input ($/1M tokens) | $2 | $5 (above 272K: $10) |
| Output ($/1M tokens) | $6 | $30 (above 272K: $45) |
| Cache read ($/1M tokens) | $0.25 | $0.50 (above 272K: $1.00) |
| Context banding | One flat rate, independent of length | Moves to a second band at 272K input tokens |
| Context window | 1,000,000 tokens | 1,000,000 tokens |
| json_object | Supported | Not supported |
| json_schema | Supported | Supported |
| Orchestration token fields | Return zero | Return above zero |
Where to go next
Read the pricing article before budgeting repeated work: on a multi-agent model the visible token count and the billed token count are not always the same, and that difference behaves differently on Fugu Max than on Fugu Ultra.
Move to the integration guide to send a first request. Keep the first experiment narrow, write your acceptance criteria down in advance, and expand volume only once measured quality and measured cost both meet your expectations.
Frequently asked questions
Is Fugu Max a single model?
No. Fugu Max is a learned orchestrator that distributes a request across a pool of expert agents. You receive one OpenAI-compatible response, but more than one model may have worked on it.
Which models are in the Fugu Max pool?
Sakana states that the pool leans on open-weights models and names the NVIDIA Nemotron family among them. The full list and the per-request routing decision are not published.
Does Fugu Max accept images?
Yes. Image input was exercised with a real PNG file on 11 September 2026 and the description the model returned matched the image. On Chat Completions you use the OpenAI-compatible image_url part.
What is the price difference between Fugu Max and Fugu Ultra?
Fugu Ultra costs 2.5 times more on input and 5 times more on output, and it moves into a second price band above 272,000 input tokens. The Fugu Max rate is a single one, independent of context length.