Product overview

Sakana AI Fugu: the model family, the API, and how billing works

Sakana AI Fugu is one name for three differently-behaving models behind a single OpenAI-compatible API. This page is the orientation: who made it, what each model is for, how to connect, what it costs, and where the official rules stop.

1. Sakana AI is the lab; Fugu is the product

Before any code, clear the naming. Sakana AI is the organization that built the family. Fugu is the product line: a set of learned multi-agent foundation models exposed through a developer console and an API. Official copy describes Fugu as coordinating a pool of models through one interface and learning collaboration patterns instead of using a hand-written workflow.

NameWhat it refers toPractical meaning
Sakana AIThe lab and vendorOwns the product page, the console, and the terms
Sakana FuguThe model familyThe product you call through the API
console.sakana.aiThe developer platformWhere Models and Pricing documentation and keys live
fugu / fugu-ultra / fugu-cyberModel IDsThe strings you send in a request

That distinction matters because Sakana AI sells more than Fugu. The same pricing page documents sakana-namazu, a separate Japanese-specialized model with its own rates. Do not fold Namazu numbers into a Fugu estimate.

2. Three models, three different operating rules

The API surface is shared, but each model has a different purpose, routing behavior, cost shape, and access requirement. Model choice is configuration, not a quality dial.

Model IDUse it forKey constraint
fuguInteractive coding, chat, normal reviews, first evaluation passPay-as-you-go cost varies with the routed model; pool opt-outs are allowed
fugu-ultraDifficult multi-step reasoning where quality is worth time and costFixed full pool; routes between one and three agents; published rates
fugu-cyberAuthorized defensive security reasoning and investigationPay-as-you-go only; returned only after access approval

Ultra currently defaults to fugu-ultra-v1.1, with fugu-ultra-v1.0 documented as the former fugu-ultra-20260615. Cyber defaults to fugu-cyber-v1.0. Call GET /v1/models rather than assuming an alias is available to every key.

3. The API is OpenAI-compatible, not OpenAI-identical

Sakana AI supports three generation surfaces plus a models list. Compatibility describes the request envelope. It does not mean every field has the same effect.

SurfacePathUse it when
Responses/v1/responsesNew integrations; recommended for tools, multimodal, and reasoning controls
Chat Completions/v1/chat/completionsAn existing OpenAI chat client cannot move yet
Messages/v1/messagesThe caller is already an Anthropic Messages client
Models/v1/modelsStartup validation of IDs and access

Two documented field rules save weeks of confusion. On Responses, temperature and parallel_tool_calls are accepted but ignored. On Chat Completions, top_p, stop, seed, the penalty fields, and temperature are accepted but ignored. previous_response_id is not accepted on Responses; send the full history in input. The full field list is on the API surfaces page.

4. Billing has a fixed part and a variable part

Ultra has published per-million-token rates. Fugu is route-dependent. Cyber no longer has a public table. Subscriptions exist beside pay-as-you-go, and orchestration tokens are extra billable usage.

Model and contextInputOutputCached input
Fugu Ultra, up to 272K context$5.00$30.00$0.50
Fugu Ultra, over 272K context$10.00$45.00$1.00
Fugu CyberNot published on August 17, 2026. Official pages say to contact sales after access is approved.

Standard Fugu pay-as-you-go is billed by route: one active agent is charged at that underlying model's rate, and several agents are charged at one rate based on the highest-tier model involved. There is no honest public fixed Fugu rate. Subscriptions are $20 (Standard), $100 (Pro), and $200 (Max) per month, described as 1×, 10×, and 20× Standard usage, with no published exact token allowance. See pricing explained for the full method.

Orchestration tokens are not a footnote.

Ultra and Cyber report orchestration input, cached input, and output fields outside the ordinary totals, and official docs say they are billed at the same category rate. A log that stores only input_tokens and output_tokens understates a multi-agent run.

5. Connect in six steps

  1. Create an API key in the Sakana console and store it in an environment variable, never in source.
  2. Set the base URL to the OpenAI-compatible endpoint and make sure it ends in /v1.
  3. Choose the model ID (fugu, fugu-ultra, or fugu-cyber) before writing the request.
  4. Call GET /v1/models with the same key and base URL and confirm the ID is present.
  5. Send the first request through the Responses API with a finite timeout and a small bounded retry count.
  6. Persist the complete usage object, including orchestration fields, for cost reconciliation.

A copyable client with startup validation and bounded retries is on the API quickstart. The order above is deliberate: a model that is missing from the list is a deployment error you should catch before paying for generation.

6. Availability and access have hard edges

  • EU and EEA: official product copy says Fugu is not yet available there while Sakana AI works toward GDPR and EU-specific compliance. Treat it as a stop, not a timeout.
  • Cyber: pay-as-you-go billing only, plus a reviewed access request describing the use case and verified contact information.
  • Partners: the product page lists OpenRouter, Vercel, opencode, Creao, and Merge. A listing is not proof that rates, Cyber access, or training opt-outs travel with that door.
  • Namazu: sakana-namazu is a separate Japanese-specialized model on the same pricing page; keep its costs separate from Fugu.

Details are split across EU availability, Cyber access, and third-party access.

7. Choosing a model, summarized

Start with Fugu for routine work. Escalate to Ultra only where a paired, blind comparison on your own tasks shows a material quality gain worth the added time and tokens. Consider Cyber only for authorized defensive security work after access is granted. If you must name every worker in an audit log, none of the managed Fugu models will satisfy that requirement, because official FAQ says the routing map is not exposed — see Fugu versus self-managed orchestration.

The model chooser encodes region, routing visibility, and workload into one recommendation, and will return "unavailable" for an EU or EEA production region before it looks at the task.

Frequently asked questions

Is Sakana Fugu free?

No permanent free tier is advertised. Subscriptions start at $20 per month and pay-as-you-go billing is available. See pricing explained.

Can I use standard OpenAI SDKs?

Yes. Point the client at the Sakana base URL with /v1 and your key. Several OpenAI-shaped fields are accepted but ignored, so validate behavior rather than assuming identical semantics.

Which model should I start with?

Fugu for interactive coding and reviews. Move to Ultra only when a blind comparison on your own hard tasks shows a real gain. Cyber is for authorized defensive security after access approval.

Where does Sakana AI end and Fugu begin?

Sakana AI is the lab and vendor; Fugu is the model family. The console at console.sakana.ai is the platform. A fuller separation is on the Sakana AI page.

Sources and verification boundary

Product facts are time sensitive and reflect the August 17, 2026 recheck. Re-read the official pages before a production decision in 2026 or later.