Sakana AI

What is Sakana AI?

Sakana AI is a Tokyo-based frontier AI research company whose usable products are two developer APIs (Sakana Fugu and Sakana Namazu), a web chat app (Sakana Chat), a translator (Sakana Translate), and a long-running research agent for businesses (Sakana Marlin).

Sakana AI was founded in 2023 by David Ha (CEO), Ren Ito (Chairman), and Llion Jones (CTO), and the company credits its team with work such as The AI Scientist, the Darwin Gödel Machine, multi-agent orchestration models, and the Namazu LLMs for Japan. Its product team describes its role as turning research into accessible tools, naming Sakana Chat, Sakana Marlin, and Sakana Fugu. A separate applied team designs and implements AI solutions for finance, defense, and intelligence work in Japan.

The latest funding round listed on the corporate page is a Series B, with Japanese backers that include MUFG, Mizuho Financial Group, Nomura Holdings, NEC, Fujitsu, and KDDI.

Core features

Fugu and Namazu are paid developer APIs; Chat, Translate, and Marlin are finished apps that reuse the same models.

Sakana Fugu: multi-agent orchestration as one model

• Learned coordination: Instead of fixed roles or hand-written workflows, Fugu is described as learning to assemble agents from a pool of models and coordinate them for each task.
• Four models, one API: Sakana Fugu comes in four models (Fugu, Fugu Ultra, Fugu Max, and Fugu Cyber), all served through one OpenAI-compatible API, so you can switch models without changing the integration.
• Fugu Ultra: coordinates a deeper pool of expert agents to maximize answer quality on hard problems, and Sakana AI says it trades response time for that depth.
• Fugu Max: orchestrates the largest pool of models and is positioned for the best cost–performance balance.
• Fugu Cyber: specialized for security analysis, vulnerability research, and threat investigation. The company reports an 86.9% success rate on CyberGym, a figure it published itself.
• Model refresh: When a new frontier model is released publicly, Sakana AI expects to spend roughly two weeks training and evaluating updated Fugu models before rolling them out.

Sakana Fugu is grounded in two ICLR 2026 papers on learned model orchestration, TRINITY and the Conductor; the product-page benchmarks are the company's own measurements.

Sakana Namazu: a Japanese LLM API

Sakana Namazu is a Japanese-specialized LLM built on the open model Kimi K2.6 and tuned with Sakana AI's in-house data for Japanese business work. It handles business documents and email in natural Japanese, including keigo and business customs, according to Sakana AI. With one request, the model itself can search the web, write and run code, and carry a multi-step task to the end.

The practical difference from Fugu is architectural: Fugu orchestrates multiple frontier models, while Namazu is a single in-house model with extra training that runs on its own as an API.

Sakana Chat

Sakana Chat launched in March 2026 as a place to try Sakana AI's models without an API, and an August update let you choose between Namazu and Fugu. The Fugu option in Sakana Chat now runs on Fugu Max.

• Code execution: the model can run Python in a sandbox for calculations, data processing, and file generation.
• Output panel: HTML renders in a side panel, where Word documents and slides can also be reviewed and downloaded.
• Attachments: images work with all models, and the models can read PDFs and Office files for summaries or analysis.
• Memory: memory works with both Namazu and Fugu Max; you can review what was stored or switch it off in settings, and it only covers conversations after the memory update.

Sakana Translate

Sakana Translate is a feature of Sakana Chat built on Namazu. It translates in both directions between Japanese, English, and Chinese and combines three modes: Translate, Proofread, and Ask. Proofread shows edits with diff highlighting and adjusts politeness, tone, and formality, not only grammar.

Sakana Marlin

Sakana Marlin is an autonomous research agent that runs sessions of up to eight hours on a topic you set, using Sakana AI's AB-MCTS method to move from framing a hypothesis through gathering information to a written report. One run produces a fully referenced report and PowerPoint slides. A later update added Interactive Reading for questioning the report and checking the evidence behind a statement, plus editable PowerPoint export.

Guide

Calling Fugu or Namazu from existing code

1. Create an account in the API Console and issue an API key.
2. Point your OpenAI SDK at the Sakana base URL and set the model to sakana-namazu, the only code change Namazu needs.
3. For Fugu, point your existing client or coding harness at the Fugu endpoint with your key; no SDK migration is required.
4. Watch the usage fields: token usage and cost are reported per request, which lets you track spend before scaling.

Trying the apps

• Sakana Chat: you can send messages without registering an email address; registering saves conversation history and raises your usage limit.
• Sakana Translate: opens in the browser with no sign-up.
• Sakana Marlin: add a card and start a run. A run takes about eight hours, and you can cancel during execution, but the credits for that run are still consumed.

Sakana AI use cases and examples

Most published examples come from Sakana AI itself and are vendor-selected.

• Hard technical work with Fugu Ultra: early users rely on it for Kaggle competitions, paper reproduction, cybersecurity analysis, and literature and patent investigations.
• Code review (company-published testimonial): a software engineer quoted by Sakana AI says Fugu Ultra surfaced more than twenty issues where other tools flagged about three.
• Recurring Japanese research reports with Namazu: a demo agent plans research, runs repeated web searches, cross-checks findings, and drafts a weekly report from one set of Japanese instructions.
• Support automation demo: another Namazu demo answers inquiry emails in polite Japanese and cross-checks order data, using fictional scenario values.
• Strategy research with Marlin: Marlin is aimed at research that combines public information, such as strategic planning, market and competitor analysis, technology trend evaluation, and regulatory shifts.

Who is it for

• Developers already on OpenAI-style APIs who want multi-agent orchestration without building their own agent framework.
• Teams whose work runs in Japanese: as a Japanese LLM API, Namazu is described as strongest on problem-solving tasks given Japanese instructions, and as worth considering for teams running frontier models at high volume that want lower costs.
• Strategy, consulting, and finance teams: Marlin is a B2B service for corporations, organizations, and sole proprietors, and is not intended for general consumer use.

It is a poor fit for users in the EU or EEA, for anyone who must see which underlying model answered, and for anyone under 18 who wants the API, which requires users to be at least 18.

Platforms

Sakana Chat currently works only from within Japan, and it recommends the latest Google Chrome on PC or smartphone. For Fugu, the company adds that access outside the EU and EEA may still be unavailable because of network conditions or local regulations. Namazu is not yet available in the EU/EEA, the UK, or Switzerland. The API terms likewise define supported regions as everywhere except the European Economic Area, the UK, and Switzerland, so the Fugu page and the API terms do not use the same list.

Fugu is also listed through third-party gateways. The OpenRouter listing for Fugu Ultra shows a 1,000,000-token context window and a maximum output of 128,000 tokens. The Vercel AI Gateway also lists Fugu Ultra, with unified billing and provider routing. For coding, Sakana AI suggests plugging the standard Fugu model into tools such as Codex for coding and code review.

Pricing

Sakana Chat is currently free, and the company says it will announce any change to pricing in advance. Sakana Translate is described as free, with no sign-up. The paid products are Fugu, Namazu, and Marlin.

Sakana Fugu subscriptions

Every subscription tier includes Fugu, Fugu Ultra, and Fugu Max. Under the API terms, the allowance does not guarantee any specific number of requests, messages, or tokens. Subscriptions renew automatically at the end of each billing period, and fees are non-refundable, including for unused time.

Pay-as-you-go API rates (per 1M tokens)

Standard Fugu has no fixed rate: when several agents are active, fees are not stacked, and you pay one rate based on the top-tier model involved. Fugu Max's web_search and web_fetch calls cost $0.007 each. Namazu charges $7.00 per 1,000 web searches and $0.12 per hour of code execution. Pay-as-you-go tokens are served at higher priority than subscription tokens. Fugu Cyber pricing is by sales contact. Namazu is billed in US dollars; its yen figures are reference conversions, and yen payment is offered only on the Enterprise plan.

API credits expire six months after purchase, and unused credits are forfeited. Independent tech press coverage of the June 2026 launch pointed out that the background tokens Fugu uses to delegate, verify, and route between agents count toward the final price at standard rates. In launch-week developer forum posts, one user wrote that $200 a month bought under three hours of use per week. Another wrote that the 5-hour limit on the $20 plan ran out surprisingly fast; these are two individual reports from before Fugu Max was released.

Sakana Marlin plans (JPY)

Every plan charges 100 credits per research run. Payment is by credit card through Stripe, and the Japanese Qualified Invoice System is supported. Plans can be cancelled or changed at any time from account settings.

Sakana AI alternatives

The comparisons below come from independent coverage and developer discussion, not from Sakana AI.

• Model routers (Not Diamond, Martian, RouteLLM): press coverage of the Fugu launch describes these as predicting which single model should answer a prompt, while Fugu breaks a query down and assigns sub-tasks to several models before combining the result.
• Agent frameworks (LangGraph, CrewAI, Microsoft AutoGen): the same coverage notes that these can build similar multi-agent systems but require manual setup of roles and state, which Fugu hides behind one endpoint.
• OpenRouter Fusion: in developer discussion of the launch, Fusion was summarized as asking several models and adding a synthesizer step, whereas Fugu uses a trained orchestrator in front of its model pool.
• Deep research assistants (ChatGPT Deep Research, Gemini Deep Research, Perplexity): coverage of the Marlin launch says these typically finish research runs in 3 to 30 minutes with 5- to 10-page summaries, while Marlin runs for up to eight hours.

Limitations

• Hidden routing: The specific models Fugu selects and how it coordinates them are proprietary, so Fugu does not show which underlying models handled a query.
• Fixed pools: Fugu Ultra and Fugu Max use fixed agent pools; only standard Fugu lets you opt specific models out from the console settings.
• Output reliability: the API terms state that Output may include false, incomplete, misleading, or outdated information and must be verified independently before use.
• Data location: for Namazu, Sakana AI cannot currently guarantee that data processing is completed entirely within Japan.
• Marlin's weak spots: Marlin is currently not suited to niche sectors with minimal public data, split-second real-time tracking, or research that relies entirely on internal, private data.
• Translate length: the Translate launch announcement cites support for up to approximately 5,000 Japanese characters. The web app checked on September 28, 2026 shows an input counter of 0/2,000, so the two pages do not match.
• Benchmarks have limits: independent coverage of the June 2026 launch reported that Fugu Ultra scored 73.7 on SWE-Bench Pro, below the 80.0 of Anthropic's limited-access Fable 5.

Privacy and data use

Training defaults differ by product:

• Fugu: you can opt out of training data use at any time from the console.
• Namazu: by default your inputs may be used to train and improve Sakana AI's models, with an opt-out in the console settings.
• Marlin: neither Sakana AI nor its external AI providers use your inputs for training unless you explicitly opt in.
• Sakana Chat: inputs may be used for training unless you opt out in settings. Conversation history and account data are stored on Google Cloud infrastructure in Japan. Limited human review may occur for safety and quality checks.
• Sakana Chat restrictions: entering your own or anyone else's personal information is prohibited. Deleting conversation history or an account is requested through the contact form.

For the API, personal data may be transferred to third parties outside your country, including to Japan and the United States. The API terms say the company is not obligated to delete model weights trained on your content, external vendor caches, or audit logs.

FAQ

Q1. Is Sakana AI one product or several?

Several. The usable products are the Fugu and Namazu APIs, the Chat and Translate web apps, and the Marlin research agent; papers such as TRINITY explain methods but are not products.

Q2. Can I use Sakana Chat outside Japan?

No, Sakana Chat is currently available only from within Japan. Fugu and Marlin work outside Japan except in EU and EEA member states.

Q3. How is Sakana Fugu billed when several models work together?

Standard Fugu charges one rate based on the top-tier model in the active pool, while Fugu Ultra and Fugu Max have fixed per-token rates. Orchestration tokens still count toward usage.

Q4. Will my data be used for training?

Fugu, Namazu, and Sakana Chat allow training use unless you opt out; Marlin uses inputs for training only with explicit opt-in.