Sakana AI's Fugu and Namazu Now Write and Run Code Inside Chat

Sakana Chat, the Tokyo-based lab's free browser interface, just received its biggest update since launch. The new version adds two models, the newly introduced Sakana Fugu and a next-generation Sakana Namazu, alongside sandboxed Python execution that lets the model produce and preview real deliverables inside the chat window.

Two models, two philosophies

Fugu and Namazu rest on fundamentally different ideas, and understanding that distinction matters for knowing which one to reach for.

Sakana Fugu is an orchestrator. Rather than a single neural network you query, Fugu acts as a conductor that bundles AIs together. Sakana AI describes it as "a full multi-agent system accessible via a single API." The design draws on two ICLR 2026 papers on learned model orchestration, TRINITY and the Conductor, which show how systems can learn to assemble, route, and coordinate expert agents for each task instead of relying on hand-designed workflows. In practice, a single Fugu request can fan out to multiple underlying frontier models playing Thinker, Worker, and Verifier roles, then synthesize a single answer back to you.

Sakana Namazu takes a different approach. It's a fine-tune of Moonshot AI's open-weight Kimi K2.6, aimed at Japanese language and Japanese business use. Sakana states plainly that it was adapted using in-house data, including tuning to reduce unnecessary refusals and limit bias in outputs. The adaptation shows up on Japanese-specific benchmarks: FairPoliticsQA (a measure of political neutrality) jumped from 34.10% on the base model to 56.30% on Namazu.

This release upgrades Namazu to a next-generation version that surpasses the previous one in both Japanese response quality and agentic execution. Fugu was previously only accessible via API, and this update brings it into the chat interface for the first time.

From talking to doing

The bigger shift here is behavioral. Sakana Chat used to return text answers. Now the model can execute Python in a sandbox, handling calculations, data processing, and file generation on its own, so you don't have to carry out the steps the model describes.

Generated outputs appear in a panel on the right side of the screen:

• HTML renders live, so you can interact with a game or app the model just built
• Word documents and slides appear as files, ready to download
• Charts and graphs from data analysis display inline
• Any output can be revised by continuing the conversation, with no re-uploading required

Image attachments work across all models, so you can paste a screenshot of an error message or a chart and ask questions about it. PDFs and Office files can also be read in for summarization or analysis.

What it's actually good for

The update leans toward Japanese-language workflows, though the code execution capability is broadly useful. Concrete use cases from the announcement:

• Vibe coding: describe a game or app in plain language and get something that runs in the browser immediately (the announcement demos a goldfish-scooping game and a kanji drill)
• Data analysis: attach an Excel file, ask for quarterly aggregation and charts; the model writes and runs the code, then outputs the graph
• Document transformation: hand it a PDF of meeting notes and ask for a slide deck; it extracts key points and rebuilds the structure
• Automated reporting: for API users, Namazu can run weekly research agents that search the web, aggregate data, and deliver a formatted report to your inbox

Fugu balances strong performance with low latency, making it a reasonable default for everyday work and a good fit for tools like Codex, chatbots, and interactive services. Fugu Ultra is tuned for maximum answer quality on hard, multi-step problems, coordinating a deeper pool of expert agents when accuracy and depth matter most.

Access and pricing

Sakana Chat is free to use from a browser with web search built in. You can try it without registering an email address, though registering saves conversation history and raises usage limits.

For API access, the two models have separate billing:

• Sakana Namazu API: pay-as-you-go with no monthly fee. $0.95 per million input tokens and $4.00 per million output tokens. Web search costs $7.00 per 1,000 calls; code execution is $0.12 per hour of session.
• Sakana Fugu API: subscription and pay-as-you-go plans, with three monthly tiers: Standard ($20/month) for lightweight daily use, Pro ($100/month) at 10x the Standard usage, and Max ($200/month) at 20x. Fugu Ultra pay-as-you-go is $5 input / $30 output per million tokens, rising to $10/$45 above 272K context.

One billing nuance worth knowing: standard Fugu pay-as-you-go is charged at the rate of whichever top-tier model is active for that request, and Sakana does not stack fees when multiple agents run simultaneously. Orchestration tokens are counted separately, though, so actual per-call cost can run higher than a simple input/output calculation suggests.

The models in Fugu's pool aren't fixed. As new frontier models arrive, Sakana expects to spend roughly two weeks training and evaluating updated Fugu models before rolling them out, meaning Fugu's ceiling should rise as the broader ecosystem does. Sakana Chat is available now at chat.sakana.ai, and the Namazu API and Fugu API are both accessible from the Sakana console.