Skip to content

Fugu Review

Sakana Fugu is Sakana AI's multi-agent system sold as one model: an OpenAI-compatible API that routes each task across a pool of frontier and open models, for developers and research teams.

from $20/mo · API

Stork's verdict on Fugu

Sakana Fugu is a serious option for hard coding and research tasks outside the EU, but expect hidden routing and variable token bills in exchange for one endpoint.

Fugu reviewed by Stork AI · stork.ai/en/fugu

stork
I am a flat drawing.
One flat drawing and one typed sentence.
This drawing was never animated.
A photo, a cartoon, a painting or a 3D character. From 10 cents a minute. Free to try.
Make a photo talkDevelopers: there is an API

What Sakana Fugu does

Sakana Fugu, from Tokyo-based Sakana AI, is a multi-agent system delivered as a single model API. Instead of you choosing a model, Fugu picks, combines and coordinates agents from a pool of models for each request. Sakana says the approach comes from two of its ICLR 2026 papers on learned orchestration, TRINITY and the Conductor. It sells four models: Fugu (the balanced default, where you can opt specific providers out of the pool), Fugu Ultra (a deeper agent pool for hard, multi-step problems, slower and pricier), Fugu Max (the largest pool, tuned for cost-performance and including NVIDIA's Nemotron family) and Fugu Cyber (security analysis, available on application only). Sakana does not reveal which underlying models answered a given request, and Fugu Ultra and Fugu Max use fixed pools.

Setup is an API key from the Sakana console and the endpoint https://api.sakana.ai, which supports the Chat Completions, Responses and Anthropic Messages formats, so existing OpenAI SDK clients work by changing the base URL. A one-line installer adds codex-fugu and claude-fugu launchers for the Codex CLI and Claude Code (Ubuntu and macOS; manual setup on Windows). Fugu is also listed on OpenRouter and the Vercel AI Gateway. It is not offered in the EU or EEA, and Sakana says access elsewhere may depend on network conditions or local rules.

Sakana Fugu pricing

Two ways to pay. Subscriptions are for everyday hands-on use and include Fugu, Fugu Ultra and Fugu Max:

PlanMonthly priceUsage
Standard$20/monthBaseline allowance
Pro$100/month10x Standard usage
Max$200/month20x Standard usage
Pay as you goUsage-basedPer-token rates below, served at higher priority than plan tokens
Fugu CyberCustomContact sales; pay-as-you-go only, access by application

Sakana does not state the Standard allowance in tokens, so the plans can only be compared with each other: Pro is $10 per Standard-sized unit of usage ($100 / 10) and Max also $10 ($200 / 20), against $20 on Standard.

Pay-as-you-go rates, USD per 1M tokens:

ModelInputCached inputOutput
Fugu Max (fugu-max-v1.0)$2$0.25$6
Fugu Ultra (fugu-ultra-v2.0), context up to 272K$5$0.50$30
Fugu Ultra, context above 272K$10$1.00$45
Fugu from Jan 5, 2027, context below 272K$5.00$0.25$25.00
Fugu from Jan 5, 2027, context 272K or above$7.00$0.35$35.00

Until January 5, 2027 (10:00 JST), the base Fugu model is billed at the standard rate of the underlying model it used; when several agents are active, you pay one rate, that of the top-tier model involved, with no stacking. Sakana does not publish those underlying rates. From January 5, 2027 Fugu moves to the fixed rates above; monthly subscription fees stay the same but allowance consumption may change. On Fugu Max, web_search and web_fetch cost $0.007 per call.

Orchestration is billed: Fugu Ultra reports orchestration input and output tokens separately in the usage object, and they are charged at the normal input and output rates on top of the visible tokens. Example on Fugu Max: 1M input tokens plus 200,000 output tokens costs $2 + (0.2 x $6) = $3.20, before any orchestration tokens. The same request on Fugu Ultra costs $5 + (0.2 x $30) = $11.00. Sakana's separate Japanese model, Sakana Namazu, is pay-as-you-go only and is not in the subscriptions.

Prices checked on the vendor's pricing page on November 10, 2026. See Sakana AI console pricing and the Fugu page.

Key features

What Sakana lists on its own pages:

  • Learned orchestration: Fugu assembles and coordinates agents from a model pool per task, with no hand-built workflow
  • One OpenAI-compatible endpoint supporting Chat Completions, Responses and Anthropic Messages formats
  • Four models: Fugu (balanced), Fugu Ultra (quality), Fugu Max (cost-performance), Fugu Cyber (security)
  • Provider opt-out on the base Fugu model through a custom model pool on the API key
  • One-line install of codex-fugu and claude-fugu launchers for the Codex CLI and Claude Code
  • Per-request token and cost reporting, including orchestration tokens
  • Opt-out of training data use from the console
  • Also available through OpenRouter and the Vercel AI Gateway

Who Sakana Fugu is best for

Fugu suits developers and research teams outside the EU who run long, hard tasks such as code review, paper reproduction, patent or literature research and agentic coding, and who would rather buy one orchestrated endpoint than build their own multi-model routing. It is also a hedge against depending on a single model vendor. It is a weaker fit if you need low latency, predictable per-request cost, or to know which model produced an answer, since routing is hidden by design; and Fugu Ultra and Max do not let you exclude providers.

If you want to choose models yourself behind one API, OpenRouter or a gateway such as Portkey or LiteLLM gives you that control. If you mainly want a single top model for coding, the Claude API is the direct route.

What users say

Independent hands-on testing is still thin; most write-ups rely on Sakana's own benchmarks and early-tester reports. Techsy, which did not say it tested Fugu itself, recommends it for research and deep code review but lists high latency, quota burn under heavy orchestration and black-box routing as drawbacks, and disputes Sakana's parity claim with one line: "Fugu Ultra is excellent, but it isn't beating the model it's compared to." Future Stack Reviews calls its piece a document review, not a hands-on test; it credits the single endpoint and the no-stacking price design but criticises hidden routing, hard-to-forecast subscription usage and the lack of a published DPA or SLA. Sakana's own page quotes unnamed early users, including a software engineer who says Fugu Ultra surfaced more than twenty issues in code review where other tools found about three; treat those as vendor-selected.

Pros

  • One OpenAI-compatible endpoint replaces building your own multi-model routing
  • Fugu Max is cheap per token at $2 input and $6 output per 1M
  • Multiple active agents are billed at one rate, not stacked
  • Ready-made launchers for the Codex CLI and Claude Code
  • Provider opt-out on the base model helps with data and compliance rules

Cons

  • Routing is hidden: you cannot see which model produced an answer
  • Not available in the EU or EEA
  • Subscription allowances are multiples of an unstated baseline, hard to forecast
  • Orchestration tokens are billed on top of visible input and output

Frequently asked questions

What is Sakana Fugu?

Sakana Fugu is Sakana AI's multi-agent system offered as one model API. For each request it selects and coordinates agents from a pool of models, so you call a single OpenAI-compatible endpoint instead of choosing models yourself. It comes as Fugu, Fugu Ultra, Fugu Max and Fugu Cyber.

How much does Sakana Fugu cost?

Subscriptions are $20/month (Standard), $100/month (Pro, 10x Standard usage) and $200/month (Max, 20x). Pay-as-you-go is per token: Fugu Max $2 input and $6 output per 1M tokens, Fugu Ultra $5 and $30 (higher above 272K context). Fugu Cyber is priced by sales. See the pricing page.

Is there a free version of Sakana Fugu?

Sakana's pricing pages list no free plan or free credits. The cheapest way in is the $20 a month Standard subscription, or pay-as-you-go by the token.

What is the difference between Fugu, Fugu Ultra and Fugu Max?

Fugu balances quality and latency and is the everyday default. Fugu Ultra coordinates more expert agents for the hardest multi-step problems, at higher cost and slower responses. Fugu Max uses Sakana's largest model pool and aims for the best cost-performance. Fugu Cyber is a separate security model available on application.

Can I see or choose which models Fugu uses?

You cannot see which models answered a request; Sakana keeps routing proprietary. On the base Fugu model you can opt specific providers out of the pool on your API key. Fugu Ultra and Fugu Max use fixed pools; enterprise customers can ask Sakana for custom configurations.

Can I use Sakana Fugu with Claude Code or Codex?

Yes. Sakana provides a one-line installer for Ubuntu and macOS that adds codex-fugu and claude-fugu launchers, with manual setup instructions for Windows. Sakana notes some mostly cosmetic compatibility differences with Claude Code.

Is Sakana Fugu available in Europe?

No. Sakana says it does not currently provide Fugu to users in the EU or EEA while it works on GDPR and other EU requirements. It is available outside Japan elsewhere, though access may be limited by network conditions or local regulations.

Compare Alternatives

Fugu

from $20/mo · API

From the Stork blog

Enjoying this? Get one like it in your inbox each morning.

one email a day · unsubscribe in two clicks · no third-party tracking