OpenAI's Browser Stopped Working Today. It Never Shipped On Windows.

The notice showed up inside the app itself, which is the part that stings. Not a blog post you could skip, not a footnote in release notes. A banner in the browser you were using, telling you the browser you were using has an expiration date. "Atlas is scheduled to stop working on August 9, 2026," it read. "Before then, export or save any important Atlas data, including bookmarks and pages you want to keep. After that date, Atlas may no longer open, browse, or support browser-based agentic workflows."

That date is today.

ChatGPT Atlas launched on October 21, 2025. OpenAI shipped it as a Chromium build with ChatGPT wired into a sidebar, free for everyone, with an agent mode reserved for Plus and Pro subscribers. The framing at launch was unambiguous. This was the company's move on the browser, the thing that stands between a person and Google.

It ran for 292 days and never left the Mac.

The Port That Never Came

Windows, iOS and Android were all named at launch as platforms Atlas would reach. None of them got a public release. By June 2026, eight months in, Atlas was still a macOS application, which means the browser built to unseat Chrome was unavailable to the overwhelming majority of desktop users on earth. Statcounter has Chrome sitting near 70 percent of global browser share. Atlas was competing for a slice of the machines Chrome does not already own, on the one platform where Safari is preinstalled and defaults are hard to dislodge.

OpenAI did keep shipping features in the meantime. Vertical tabs landed in November 2025. Tab groups landed in January 2026. Browser memories, the feature that let Atlas recall what you had been looking at, stored that context on OpenAI servers for 30 days. These are the additions of a team that intended to stay. They are also, in retrospect, the additions of a team polishing a hull while the shipyard was being rezoned.

In March 2026 OpenAI announced it would merge Atlas with the ChatGPT desktop app and Codex into a single desktop application. That was the tell. A standalone browser does not get folded into a chat client unless somebody has decided the browser was the wrong container.

What Actually Killed It

On July 9, 2026, James Sun, who leads OpenAI's browsing work, posted the announcement on X alongside the launch of ChatGPT Work. "Finally, with all these updates, we are going to be sunsetting Atlas," he wrote. "All these capabilities were built on what we learned from Atlas users who took a leap of faith on a new browser."

Leap of faith is doing heavy lifting in that sentence.

The replacement is three things instead of one: ChatGPT Work, which fuses ChatGPT and Codex; an upgraded browser inside the desktop app, plus a cloud browser that runs agent tasks on OpenAI's own servers; and a Chrome extension that reads the page you are on and can answer questions, summarize it, or start a longer task. The strategic logic is defensible and OpenAI states it plainly enough: it is easier to put AI browsing inside the browser people already use than to convince them to install a different one.

The uncomfortable version is that the product was slow and unsafe in ways nobody solved. The Verge, testing agent mode, watched it take ten full minutes to add three items to an Amazon shopping cart. Ten minutes. A human does that in ninety seconds and does not need a subscription tier.

Security was worse. In October 2025, within days of launch, LayerX Security published a cross-site request forgery technique it called "ChatGPT Tainted Memories," which chained into prompt injection. OpenAI disputed that the finding reproduced, and no exploitation in the wild was ever documented. But the category never went away. By December 2025 OpenAI was publishing that prompt injection is "one of the most significant risks we actively defend against to help ensure ChatGPT Atlas can operate securely on your behalf," while conceding it may never be fully solved for browser agents, an admission that lines up with the UK National Cyber Security Centre's earlier warning about the same class of attack.

An agent that browses on your behalf has to read untrusted text and then decide what to do. Those two jobs are in permanent conflict. OpenAI did not fix that. Nobody has.

The Side Quest Problem

Atlas did not die alone. It died during a company-wide reckoning about how many things OpenAI was building at once. Fidji Simo, then CEO of Applications, told staff in an all-hands that Anthropic's enterprise traction was a "wake-up call" and that OpenAI "can't miss the moment because we are distracted by side quests." Kevin Weil and Bill Peebles both exited in April 2026 as the company shed projects. Simo herself stepped down from the number two role on July 9, 2026, the same day Atlas was sunset, citing a neuroimmune condition whose recovery "proved longer and harder than expected," moving to a part-time advisory seat. Sam Altman's public response was short: "i am really sad about this and very grateful for all fidji has done for openai."

Read the calendar and the shape is clear. A browser announced as a Google-killer became, inside twenty-one months, a line item in a portfolio cleanup.

What Users Have To Do Today

Bookmarks do not transfer automatically. Anyone who has been using Atlas as a daily driver has to export them as an HTML file and import them somewhere else by hand. Conversation history is stored separately from the browser and survives. Cookies may be exportable depending on what you were doing. The app itself will stop receiving security updates, which turns any copy left installed on a Mac into an unpatched Chromium build with network access, and that is a genuinely bad object to keep around.

Workspace admins were told to notify their members. Some of them will not have.

The Honest Counterargument

Killing Atlas was probably the right call. A browser is an enormous, permanent maintenance obligation: four platforms, a rendering engine you inherit from Google, a security update treadmill that never stops, and an install-base problem that has defeated far better funded companies than OpenAI. If the agentic capability is the actual product, then shipping it as an extension into a browser with 70 percent share reaches more people on day one than a macOS-only browser reached in nine months. Cutting a losing bet quickly is a competence, not a scandal.

The scandal is the gap between how it was sold and how it ended. Atlas was launched with the rhetoric of a platform, a challenge to the default way people find information. Platforms come with an implied contract that they will still exist in two years. This one did not last one, and the users who moved their bookmarks and their browsing habits into it are the ones spending today doing an export they were given a month's notice to complete.

Perplexity shipped Comet just before Atlas arrived. Anil Dash described Atlas as an "anti-web browser," a thing that fed you summaries instead of pages. Publishers worried about losing referral traffic to a product that answered questions without sending anyone anywhere. Those criticisms all assumed Atlas would matter. Its actual legacy is smaller and stranger: proof that OpenAI can conceive, build, ship, market, and bury a flagship consumer application inside a single fiscal year, and that a banner in the app is how you find out.