ChatGPT Atlas Shuts Down August 9. Read the Shutdown Notice, Not the Launch Post.

OpenAI's AI browser lasted under ten months. Its migration guide lists exactly what a browser owns that the app replacing it does not, and that list is the audit worth running on every agentic tool you use.

ChatGPT Atlas launched on Mac in October 2025. It stops working on August 9, 2026. OpenAI's support article gives users what it calls "an approx. 30 day wind down period," measured from a July 9 announcement, and inside that article is a sentence most product deprecations never print: "Browsers require ongoing security maintenance, and we do not want users to remain on a discontinued browser that may degrade or stop receiving security updates."

A browser that will stop getting security patches on a known date is a specific kind of dead. Five days from now, anyone still running Atlas is running an unmaintained browser with live cookies in it.

The launch post told you what Atlas was for. The shutdown notice tells you what it was holding.

The AI browser was a distribution bet, and the bet closed

Here is my position, and it is not a eulogy. The standalone AI browser was never its own product category. It was a way to get an agent close to the page, and once the desktop app could do multiple tabs, downloads, navigation, and account logins, the browser stopped being the only door. OpenAI says as much: it is "deprecating Atlas and moving browser-based agentic capabilities into ChatGPT and Codex," and building on what it learned from Atlas to support a more capable browser experience inside ChatGPT itself.

Consolidation is a defensible engineering call. Maintaining a browser is a permanent security treadmill, and OpenAI decided not to run it for a surface it was folding into another product anyway.

What I care about is the bill. Every consolidation has one, and it is always paid in state that has no export path. Atlas's notice is unusually specific about where that state sits, which is why it is worth reading even if you never installed the thing.

What the notice says a browser owns

Work through the FAQ and a list falls out of it.

Bookmarks do not transfer automatically. You export them to an HTML file yourself, then import that file into another browser.

Open tabs "may not transfer automatically." OpenAI's suggested workaround is to bookmark the pages or copy the URLs into a document by hand.

Browser history will not transfer automatically. Same fix, done manually, before the date.

Cookies get the most careful wording on the page. Atlas "will provide export options where available," and OpenAI adds a warning I wish more vendors wrote: "Do not share cookie or session files unless you trust the recipient and understand what access the files may provide." A separate FAQ answer says Chrome may not natively import cookies at all unless you have a supported import path. So the one piece of state that represents your logged-in identity across every site you touched is the piece with the weakest export story.

ChatGPT conversation history is the exception. It lives in ChatGPT, not in Atlas, and it survives.

Then there is the admin section, and it contains the line I keep thinking about: "Workspace admins should review whether members are using Atlas and share this article with affected users." Read that as an instruction and it is reasonable. Read it as a description of the tooling and it says something else. The admin is being asked to go find out who is using the product, which implies the console does not simply tell them.

The same consolidation is happening one layer down

Atlas is the visible half. The other half showed up in the Codex changelog, and it is worth naming because it lands three weeks after the browser does.

On August 31, 2026, GPT-5.4 and GPT-5.4 mini stop being available in Codex for users signed in with ChatGPT. The changelog points those users at GPT-5.6 Terra and GPT-5.6 Luna instead. The same two models stay available on the OpenAI API and in Codex sessions authenticated with an API key.

So the model you can reach depends on which credential you signed in with. That is not a hypothetical. It has been live and surprising since spring: issue #14412 on the openai/codex repo was opened on March 11 by a user who found that models working on March 9 returned {"detail":"The 'gpt-5.4' model is not supported when using Codex with a ChatGPT account."} on March 10. The maintainers labeled it auth and bug. The reporter noted the same behavior across Free and Business plans.

Two deprecations, three weeks apart, both consolidating a surface. One removes a product. The other removes a model from one door and leaves it open on another.

Put this into practice

None of this takes long. The Atlas half has a deadline five days out.

Export bookmarks today, not on the 8th. Open Atlas, open the bookmark manager, choose export, save the HTML somewhere you will find it. In Chrome: More, then Bookmarks and lists, then Import bookmarks and settings, then pick the file. Ten minutes.
Copy your open tabs into a plain text file. Not because it is elegant, but because "may not transfer automatically" is the kind of phrasing that resolves badly on the day.
Treat any exported cookie file as a secret. If you export one, move it, and delete it. Do not put it in a shared drive, a ticket, or a chat message. A session file is a live credential wearing a filename.
Grep for Atlas in your own automation. Shortcuts, launch agents, shell aliases, anything that opens a URL in a named browser. A deprecation only hurts when something you forgot about was pointed at it.
Do the auth audit before August 31. On every machine and every CI job running Codex, write down two things: which auth path it uses (ChatGPT sign-in or API key) and which model it pins. If a ChatGPT-authenticated session pins gpt-5.4, swap it to gpt-5.6-terra. If it pins gpt-5.4-mini, swap it to gpt-5.6-luna. If you need those exact models past the end of the month, that session moves to an API key.
Build the export column for whatever you adopt next. Two columns on one page. Left: every kind of state this surface holds (files, credentials, sessions, history, memory, configuration). Right: the export format for each. Any row where the right column is empty is your lock-in, written down before you are attached to it. That page takes fifteen minutes and it is the only part of this that keeps paying after August.

Honest limitations

Thirty days is short, and I said so. It is also more than a lot of deprecations offer, and OpenAI published a dated article with named steps, an admin section, and troubleshooting for the two things most likely to go wrong. That is better handling than the median shutdown, and pretending otherwise would be cheap.

The hedged phrasing cuts both ways. "May not transfer automatically" might mean the real behavior is better than the doc promises. I have not run the export on a current Atlas build, so every step above comes from OpenAI's article rather than from my own migration.

The consolidation argument deserves its strongest form. Running a browser means shipping security updates forever, for a surface that duplicates capability the desktop app now has. Killing it is the responsible version of that decision, not the cynical one, and the alternative (a slowly rotting second browser) is worse for users than a hard date.

On the Codex model change, I want to be precise about sourcing. The developers.openai.com changelog page I fetched on August 4 served entries only through June, so I did not read the August 31 entry on the page itself. The dates, the auth split, and the two replacement model IDs are consistent across search results and match the behavior already documented in issue #14412, but treat them as reported rather than as something I loaded end to end. Check your own changelog before you rewrite a config on my say-so.

I also cannot tell you whether the ChatGPT desktop app's browsing is a fair replacement for Atlas in daily use. OpenAI lists multiple tabs, downloads, improved navigation, and account login support "where available," and that last phrase is doing work. Availability depends on plan, region, device, and workspace settings.

What this actually asks of you

The useful habit here is smaller than the news. Every agentic surface you adopt accumulates state, and the vendor's own migration document is the only place that state gets enumerated honestly, because writing it is a support obligation rather than a marketing choice.

So pick one tool you rely on right now. An agent harness, a browser extension, a hosted sandbox, a memory service. Ask the boring question: if this had a shutdown notice next month, what would be on it, and which line would have no export format?

If you cannot answer that in a couple of minutes, you have found the thing worth fixing this week. And if you already ran the Atlas export and something did not come across the way the doc said it would, I want to hear which piece it was.

Sources: OpenAI Help Center, "Evolving Atlas into ChatGPT for browser-based agentic work"; openai/codex issue #14412; Codex changelog; Search Engine Land, "OpenAI sets Aug. 9 end date for ChatGPT Atlas".