Sakana AI updated Sakana Chat on August 13, 2026: Fugu is now a selectable model, Namazu moved to a new generation, and the chat window can run Python in a sandbox, preview the artifact, and take image and document attachments. The Japanese launch thread framed it as vibe coding — games and apps from words alone, from kanji drills to goldfish scooping.
That is a product update, not a new weight drop. Fugu is the multi-model orchestrator we already covered. Namazu was already the Japanese-first model. The delta is the surface: no API key, code that actually executes, and files you can drop in.
TL;DR — what people are actually asking
| Question | Direct answer |
|---|---|
| What shipped? | Fugu + new Namazu in Sakana Chat, plus Python sandbox, artifact panel, attachments |
| When? | August 13, 2026 (Sakana's JP/EN post and X thread) |
| New Fugu weights? | No. Chat access to the existing orchestrator family |
| New Namazu? | Yes — a generation bump; no new public eval table |
| Code execution? | Sandboxed Python; HTML/Word/slides preview on the right |
| Attachments? | Images on all models; PDFs and Office docs for summary/analysis |
| API required? | No for Chat. Namazu API still exists separately |
| Who is this for? | Japanese-language vibe coding and office loops, not a Cursor replacement |
What Sakana Chat shipped on August 13
Sakana's own English post lists two jobs for the update:
- More models to pick. Fugu joins Chat. Namazu is refreshed.
- The model does the work. Code execution plus an artifact panel, so Chat no longer ends at a text answer.
Attachments are the third piece. Images work across all models. PDFs and Office files can be summarized or analyzed. Combined with the sandbox, the intended loop is: drop a file, ask in Japanese, get a chart, a slide, or a playable HTML page back in the same thread.
Chat remains the no-API try-it surface. Namazu is still on the API. Fugu's paid path is still the Claude Code-compatible launcher and the API — not this browser window.
No prices, rate limits, or region restrictions landed in the Chat post. Do not inherit the July Fugu API notes (EU/EEA unavailable, training opt-out in the console) onto Chat unless Sakana says so. Different product.
Fugu and Namazu now share one chat window
Fugu is the "quality of the response determines the quality of the outcome" pick: complex instructions, multi-step work. That matches how Sakana has sold the orchestrator since June — a coordinator in front of a pool, not a single dense model you inspect. Inside Chat you still will not see which specialists ran. Treat it as the expensive, ambitious default, the same way you would pick Fugu-Ultra in a terminal.
Namazu is the Japanese daily driver. Sakana's claim for the new generation is twofold: better Japanese text, and better agentic execution. The product sentence that matters: it is built to take Japanese instructions as they are and return natural Japanese documents and materials. That is the opposite of "prompt in English, translate the output."
There is still no Namazu eval card in this announcement. No AA Index, no Terminal-Bench, no Harbor shader time. If you need numbers, you are still on the older Fugu Harbor / orchestration write-up and the Fugu-Cyber benches — different products.
| Surface | Model | What it is for |
|---|---|---|
| Sakana Chat | Fugu | Hard, multi-step work; frontier-adjacent orchestration |
| Sakana Chat | Namazu (new gen) | Japanese documents, worksheets, office loops |
| API | Namazu | Same family, programmatic |
| API / Claude Code launcher | Fugu / Fugu-Ultra | Builder harness, paid, inspect-nothing routing |
If you already run Fugu in Claude Code, this Chat update does not change your terminal setup. It gives a colleague who will not install a launcher a place to try the same lab's models.
Python sandbox, artifacts in the side panel
Until this week, Sakana Chat stopped at prose. The model can now execute Python in a sandbox: calculation, data processing, file generation. You are not supposed to copy the steps out and run them yourself.
Outputs land in a right-hand panel:
- HTML renders as it would in a browser — including simple games you can play in place.
- Word and slides preview as documents.
- You can download the artifact, or keep talking and have it rebuilt in the same slot.
The Excel example in the official post is the office version of the same loop: attach a spreadsheet, ask for quarterly aggregation and a chart, and the model writes the code, runs it, and returns the graph. Research → compute → deck, one thread, no tool-hopping.
That is also the security boundary to notice. A chat product that executes model-written Python and accepts Office/PDF uploads is a different threat model than a chatbot. Sakana says "sandbox." It does not publish sandbox limits, network egress, or retention. Do not drop confidential workbooks in a free consumer chat until you have read the live terms. Same advice we give for any in-product code runner.
Attachments are the Japanese office loop
Image attachments are on all models: paste an error screenshot or a chart and ask about it. Documents (PDF, Office) can be summarized or analyzed. The worked example is a meeting-PDF → distilled slides.
This is closer to "Claude in the browser with a Python sidecar" than to Cursor. The launch demos Sakana posted on X are consistent with that: math worksheets, kanji drills, goldfish scooping, Excel-to-graph. Kids on summer break and Japanese business users are the named audiences. Software engineers with a repo are not.
If your job is a long-running coding agent, stay on Grok Build, Cursor, or the Fugu Claude Code path. Chat is for people who will not open a terminal.
Vibe coding that is not English-first
Sakana's X copy uses the English loanword Vibe coding and then shows Japanese cultural tasks. That is the point. Most vibe-coding demos — stadium shaders, landing pages, clone-the-screenshot — are English-prompt, English-UI, English-docs. Quality falls off as you leave English, which is why a reply under the launch thread noted that tuning a local model beats white-labeling a US one.
We have already defined the term in the vibe coding explainer: demo-speed generation, often without tests. Sakana Chat's new loop (prompt → sandbox → playable HTML) is exactly that shape. The interesting part is language, not a new coding paradigm.
What to verify yourself, if you read Japanese:
- A worksheet or drill that a child can actually use, not a pretty PDF that fails on the second page of problems.
- An Excel attach → chart → slide restyle on a real internal-looking (non-secret) spreadsheet.
- A small game (the goldfish-scooping demo class) and whether iterate-in-place actually rebuilds the panel or silently forks a new file.
What not to expect: a Terminal-Bench number, a CursorBench row, or Fable-level English repo edits. Those are other products from the same lab.
What Sakana Chat does not replace
- Fugu API / Claude Code launcher. Chat is not a substitute for
ANTHROPIC_BASE_URLpointed at Sakana. Builders still want the setup guide. - A published Namazu leaderboard. Generation bump, no card.
- English-first agentic IDEs. Cursor, Claude Code, Grok 4.6 in-IDE are still the default for English repos. See this week's Grok 4.6 eval card.
- Sakana's research line. Diffusing Blame and CoffeeBench are a different audience than Chat.
The honest one-line: Sakana put its orchestrator and its Japanese model into a browser, then gave them a Python sandbox so "make this" can mean "here, play it / download it." That is a product completion for Chat, not a frontier-model launch.
Related on explainx.ai
- Sakana Fugu — one API that orchestrates the others
- Sakana Fugu in Claude Code: setup, pricing, limits
- Fugu-Cyber — orchestrated cyber defense benches
- What is vibe coding?
- Grok 4.6 launch evals
- What is an agent harness?
- Loop engineering for coding agents
- What are agent skills?
- Asian AI alternatives after the Mythos export gap
Primary sources: Sakana AI, Sakana Chat update (August 13, 2026) · Sakana Namazu product page · Sakana AI on X (August 13, 2026)
Product details are accurate as of August 13, 2026 and reflect Sakana's Chat update post. No new public Namazu or Fugu-in-Chat eval numbers were published with this launch. Sandbox limits and data retention are not specified in the announcement — read the live Chat terms before uploading work files. Follow @explainx_ai for updates.