Atlas is deprecated August 9, 2026, with bookmarks, history and open tabs requiring manual export. The strategic read: agentic browsing is converging on a protocol layer inside existing browsers — Microsoft folded Copilot into Edge, and WebMCP entered an origin trial in Chrome 149 — rather than on a new browser.
ChatGPT Atlas stops working on August 9, 2026, less than a year after OpenAI launched it. If you used it, you have days to export your data, because most of it does not move on its own. That is the urgent part. The more durable part is what the retreat says about where agentic browsing actually lives, because OpenAI is replacing a browser it controlled with an extension inside a browser it does not.
Per OpenAI's own help documentation, Atlas data including bookmarks, browsing history, open tabs, saved passwords and cookies will not transfer automatically to another application. Reporting indicates cookies and saved passwords will carry into the ChatGPT desktop app, but bookmarks, history and open tabs are your responsibility.
After August 9, access ends permanently. OpenAI has not described a grace period for retrieving data from a deprecated Atlas install.
Atlas capabilities move to two places: an upgraded browser built into the ChatGPT desktop app, with multi-tab browsing, website logins, downloads, autofill and password manager integration, and a new ChatGPT extension for Chrome that can read page context.
The case for building a browser rather than an extension was never about rendering web pages. It was about control of the layer where an agent operates. A browser you own gives you persistent authenticated sessions, unrestricted access to the DOM, cross-tab state, the ability to design an interface around agent actions rather than bolt it onto someone else's, and no dependency on a competitor's extension policies.
That is a real list. An extension operates inside a sandbox that Google defines, subject to permissions Google reviews, in a product Google can change. For a company whose stated ambition is agents that act on your behalf across the web, owning that surface looked like the defensible position.
Against that control, a new browser starts with zero distribution and a trust problem it cannot engineer around. Browser switching is one of the stickiest behaviours in consumer software. Users have to be persuaded to move their bookmarks, their extensions, their saved logins and their muscle memory, and then to trust a months-old application with every password they own.
Chrome already has the users, the credentials and the habit. If the differentiating asset is the model rather than the rendering engine, then shipping the model where the users already are dominates shipping a browser and asking them to move. The extension is not a smaller version of the Atlas strategy. It is the opposite strategy, and it concedes that the browser was never the moat.
“A new browser has to win a trust argument before it can start the product argument. An extension inherits the trust and starts at the product.”
Worth stating plainly: this is a reading of a decision, not an explanation OpenAI has given. The company frames the change as evolving Atlas into ChatGPT for browser-based agentic work rather than as a retreat, and it is possible to run a strategy in which a standalone browser was always a capability testbed. The observable fact is the direction of travel, from a controlled surface to a borrowed one.
Microsoft retired the Copilot Mode brand and folded AI capability directly into Edge rather than maintaining a separate agentic surface, according to industry analysis. Two of the largest players in the category have now concluded, independently, that agentic browsing belongs inside a general-purpose browser rather than beside one.
The more consequential development is standards work. WebMCP, a proposal for letting websites expose structured tools to browser-based AI agents rather than forcing agents to interpret a page visually or by scraping the DOM, entered an experimental origin trial in Chrome 149. If it matures, an agent would call a declared function on a site instead of simulating a click on a button it has to identify from pixels.
That changes the calculus completely. Most of the advantage of owning the browser comes from needing privileged, unmediated access to page internals because the page will not cooperate. A protocol in which the page does cooperate removes the need for the privilege. The agent does not need to own the browser. It needs the browser to speak a protocol.
An origin trial is an early, reversible experiment, not a shipped standard. WebMCP may not survive contact with adoption. Sites have to implement it, and sites with a commercial interest in blocking agents have an obvious reason not to.
If the protocol reading is right, the competitive question in agentic browsing shifts from who ships the best browser to who defines the interface between agents and sites. Google, which controls both the dominant browser and the origin trial, starts that contest from an unusually strong position. The counter-scenario is that WebMCP stalls, page cooperation never materialises at scale, and controlling the runtime becomes valuable again for exactly the reasons Atlas was built.
Three things will indicate which is happening within a year. Whether major sites, not demos, ship WebMCP endpoints. Whether OpenAI's Chrome extension hits a permissions wall that a browser would not have. And whether anyone launches a new agentic browser after this. If the answer to the last one is nobody, the category has resolved.
For now, the actionable part remains the first section. Export your bookmarks. Atlas is gone Friday, and OpenAI has been explicit that it will not carry your data across for you.
Export bookmarks from Atlas as an HTML file and import that file into your next browser. OpenAI's guidance is that Atlas data, including bookmarks, history, open tabs, saved passwords and cookies, does not transfer automatically to other applications, though reporting indicates cookies and passwords carry into the ChatGPT desktop app. Verify your credentials in the destination before August 9, 2026, after which access ends permanently.
DrafterDaily covers the infrastructure shifts behind the product news.
Read more Technology analysisCVE-2026-0768 is an unauthenticated root RCE in Langflow. It was disclosed in January, the fix has shipped through seven releases, and attackers are hitting it in September — because low-code AI middleware became critical infrastructure without acquiring a patch owner.
Opposition to data centres is not a referendum on AI. It is a cost-allocation dispute — and PJM's capacity auction is the mechanism that turned an abstract argument into a number on 67 million households' bills.
OpenAI's first published benchmarks for its custom inference chip are real, from a public benchmark, and normalised on nameplate TDP rather than measured power. OpenAI disclosed that itself, in a sentence nobody is quoting.