The deadline is Saturday, August 9, 2026. OpenAI is retiring the standalone Atlas browser that day — no automatic migration, no grace period, no "we'll import it for you." If you built anything on Atlas, an agent workflow or just a folder of bookmarks, the useful version of this article is one paragraph long: export your data today, and decide where the agentic work goes this weekend. Everything below is how.

The 10-minute export, in order#

Atlas does not hand your data to its successor. Bookmarks, saved passwords, and local browsing data do not carry over on their own. Do this before the 9th:

  1. Bookmarks. Export them from Atlas as an HTML file. Then import into whatever browser you're landing on — in Chrome that's More → Bookmarks and lists → Import bookmarks and settings, and point it at the HTML file. This is the one that quietly disappears if you skip it.
  2. Passwords. If you let Atlas store logins, export them (or confirm they already live in your password manager / Google account) before access ends. After the 9th there is no "open Atlas real quick to grab that password."
  3. **What you do not need to panic about: your ChatGPT conversation history**. It's stored in your ChatGPT account, not inside Atlas, so it's unaffected and still there in ChatGPT after the shutdown. That's the one piece OpenAI is not asking you to save.

That's the whole backup. Ten minutes, mostly the bookmarks step.

Atlas launched in October 2025 and is being switched off in August 2026 — a nine-month lifespan for a shipped browser. That's the part worth sitting with before you pick a replacement.

Where the agentic browsing actually goes#

OpenAI isn't killing the capability — it's moving it. The autonomous browsing that made Atlas interesting (driving multiple tabs, downloading files, navigating, signing in to accounts and acting on your behalf) is being folded into the new ChatGPT desktop app and a Chrome extension. The bet OpenAI is making out loud: browsing is a feature of ChatGPT, not a product you install.

For a founder who was already all-in on ChatGPT, that's the lowest-friction move — the agent follows you into an app you already open. Install the desktop app, sign in, and your agentic browsing lives inside the assistant instead of a dedicated window. No migration of workflows, because the workflow was a prompt, not a config file.

But the shutdown quietly reopens a bigger question for everyone who built real automation on Atlas.

The real decision: which agentic browser now?#

If Atlas was just where you read things, the ChatGPT desktop app is fine and you can stop reading. If Atlas was doing work — scraping, filling forms, running logged-in flows — you now get to re-pick your agentic browser from scratch, and the field in August 2026 is genuinely competitive.

The honest cut: stay on ChatGPT if you value zero friction and already pay OpenAI. Move to Comet if you live in research and want an actual browser that's already on every platform. Reach for Claude in Chrome + Cowork if you're a builder who already pays Anthropic and wants coding-grade agents on the page — we compared that agent head-to-head in Claude Cowork vs ChatGPT Work. And if the thing that stings about this whole episode is that a vendor can switch off your browser with a month's notice, the answer is BrowserOS — an open-source agentic browser you run yourself, with your own model behind it. It can't be sunset out from under you because there's no "you" to sunset.

The full field, side by side:

(see the comparison table above)

What it means#

Two things, and neither is "cry about Atlas."

First, do the export today, not Friday night. August 9 is a hard wall — the kind of dependency deadline we keep putting on founders' calendars, right next to the Atlas-and-Assistants-API drill in this week's Wire. A ten-minute task becomes a permanent data loss the moment the wall arrives.

Second, treat the destination as an architecture choice, not a brand loyalty test. The agentic-browser layer is moving fast enough that betting your automation on any single vendor's standalone app is how you end up doing this migration again in nine months. If the workflow matters, either keep it model-agnostic (the same argument we made about running a browser engine you control) or run something open you own outright. Atlas's nine-month life is the cheapest reminder you'll get that "the browser is the agent" cuts both ways: when the vendor moves on, so does your agent.

Export first. Decide second. The 9th doesn't wait.