Sakana AI
Sakana AI · Other
Sakana AI is a Tokyo research lab that sells frontier models through model merging. Its lineup includes the Fugu multi-agent AI system, the Japanese LLM Namazu, and the Marlin research agent.
About Sakana AI
What Is Sakana AI
Sakana AI is a Tokyo company that builds AI systems by combining existing models instead of training one huge model from scratch. It was founded in July 2023 by David Ha, Llion Jones and Ren Ito. Jones co-authored the 2017 paper that introduced the Transformer, the architecture behind most modern language models. The name means fish in Japanese, and the company's research borrows ideas from schools of fish and other natural systems.
The company sells model access through an API rather than a consumer app. Its flagship product, Sakana Fugu, sends a request to a pool of frontier models and lets a trained coordinator decide which ones handle which part of the job. Sakana Namazu is tuned for Japanese business writing. Sakana Marlin writes long strategy reports for analysts. Most buyers are companies with engineering or research teams, not individual users.
Geography is the first limit you hit. Fugu is not available in the EU or EEA, and Namazu is also blocked in the UK and Switzerland while the company works toward GDPR compliance. There's no free tier and no chat app, so you need a developer or a vendor to put the models to work.
Getting Started
- Create an account in the Sakana API console. A work email and a company name are enough to start.
- Issue an API key. Fugu and Namazu keys are managed separately, so decide which product you're testing first.
- Point an OpenAI-compatible client at the Sakana base URL and set the model name, such as
sakana-namazuorfugu. - Send a short test prompt, then compare latency and token counts against the model you use today.
- Set a spending cap in the console and connect the key to the tool your team already runs.
Product Information
A quick look at Sakana AI's pricing, supported platforms, and performance.
Best for
The users, tasks, and scenarios where this tool fits best.
Users
- Engineering teams with long agentic jobs
- Product teams that serve Japanese customers
- Strategy and research groups
Tasks
- Multi-step coding and code review
- Japanese document work
- Early-stage market or technical research
Scenarios
- You want one contract instead of five model vendor accounts
- A provider goes down or changes terms
- A regulated team needs to keep certain models out
Key features
One Endpoint for a Pool of Frontier Models
Fugu looks like a normal chat completion call. Underneath, a coordinator model decides whether a request needs one model or several, assigns roles, and merges the results into a single answer. You don't manage the routing, the retries or the intermediate prompts.
Fugu Ultra for the Hardest Tasks
Sakana ships several Fugu variants. Fugu balances speed and quality for everyday coding and chat. Fugu Ultra spends more time and more tokens chasing the best answer, and the company positions it against top-tier reasoning models on software engineering and science benchmarks. Fugu Cyber is tuned for security work, and Fugu Max targets cost against performance.
Model Pool Controls
You can remove providers or individual models from the pool. That matters for companies with data residency rules, procurement bans or a policy against sending code to a specific vendor. The setting lives at the account level, so one team can't quietly re-enable a blocked model.
Namazu for Japanese Business Writing
Namazu is built on the open Kimi K2.6 model and tuned with in-house Japanese data. It writes business email, internal documents and customer replies in natural keigo, and it can search the web and run code mid-task. Switching is a one-line change if your stack already speaks the OpenAI SDK.
Marlin for Analyst-Grade Research
Marlin is an autonomous research agent rather than a chat model. Give it a question about a market, a competitor or a technology, and it reasons for an extended period, gathers sources and returns a strategy report written for consultants, corporate strategy units and think tanks. Sakana sells it to firms that already pay analysts for the same work.
Research You Can Read
The products sit on published work. TRINITY and the Conductor, both ICLR 2026 papers, describe how a learned coordinator can assign Thinker, Worker and Verifier roles across a model pool. The AI Scientist, built with Oxford and UBC, became the first fully automated research system to pass peer review at a machine learning workshop, and the code is on GitHub.
Pros and cons
Pros
- One API key reaches several frontier models, which cuts procurement and integration work.
- Billing follows the most expensive model used in a request instead of summing every model, so multi-agent runs stay predictable.
- Japanese output is a real strength, not a translated afterthought.
- The research behind the products is public, so you can judge the method before you buy.
Cons
- No access in the EU or EEA, and Namazu is also blocked in the UK and Switzerland.
- No free tier. You need a paid key before you can test anything.
- The coordinator is closed and depends on third-party models, so quality can shift when a provider updates its own model.
- Documentation and support are thinner than what larger vendors offer, and the console assumes you're comfortable with API tooling.
Frequently asked questions
Sakana AI sells API access to models it builds and tunes: Fugu for multi-agent orchestration, Namazu for Japanese business work, Marlin for research reports and Sakana Translate for Japanese, English and Chinese translation. There's no consumer subscription.
Related content
Explore related tools, skills, and articles for Sakana AI.
Sakana AI Alternatives
Odyssey 3
Odyssey · OtherOdyssey-3 is Odyssey's foundation world model. It simulates physics in real time, turns a prompt into an interactive environment, and adapts to robots, cars and games.
JarvisCore
Prescott Data · Coding · Productivity · OtherJarvisCore is an open source Python runtime where AI agents work as equal peers on an agent mesh, claim tasks from a shared ledger, and reach external services through a zero-trust broker.
Clef
Cloudflare · Business · OtherClef is Cloudflare's open-source decision model for Workers AI. It returns calibrated probabilities for typed questions, not generated text, so agents can classify, route, and act fast.