OpenAI drops its own browser and moves its agent into Google's Chrome, a sign that distribution matters more than owning the browser

ChatGPT can now see the Chrome tabs you already have open. OpenAI has upgraded its Google Chrome extension, The Verge reports. If you tag Chrome in ChatGPT's web or desktop app, the assistant can pull information from those tabs. This comes after OpenAI discontinued ChatGPT Atlas, its own AI browser. The message is plain: OpenAI no longer needs you to switch browsers. It is going to the browser you already use.

OpenAI's demo shows a user asking ChatGPT to compare three open tabs, each showing a sofa they are considering. The company's sample prompts go further. One asks ChatGPT to copy vendor details from several tabs into a comparison sheet with prices, key features and source links. Another asks it to compare contractor services and draft an email to the best fit. A third asks it to fill in a new dentist's intake form and add the appointment to both a work and a personal calendar. The extension also puts ChatGPT in Chrome's sidebar with what OpenAI calls "the full power of ChatGPT" and your full conversation history in Chat or Work. The biggest step is that ChatGPT Work can now "control" Chrome tabs. It can send messages, fill out forms, download files, and organize or group tabs. The Verge compares this to Google's Gemini-powered "auto browse" feature.

The context explains why this matters. We have argued that the AI race is moving from model quality to the "plumbing": who controls distribution, integration and the standards agents run on. Atlas was OpenAI's attempt to own that plumbing outright. Getting people to leave the browser they already use is very hard. Retiring Atlas and putting the agent into Chrome looks like a pragmatic retreat. OpenAI is choosing reach over control. The cost is that it now operates inside a platform run by Google, a direct rival with its own built-in agent. If Google changes the extension rules, OpenAI has to adapt.

There is also a short-term risk that the launch materials don't address. An agent that can read and act in your logged-in tabs can reach your email, your accounts and your forms. One of OpenAI's own examples says "stop before submitting." That is a hint that keeping a human in the loop is still the safety net. A related risk, in our assessment, is that the open web is full of content an agent could misread as instructions. The case for this feature rests on how well it is contained: clear permissions, confirmation before sending or paying, and logs of what the agent did. We have said before that when you can't fully limit how an agent behaves, you have to limit the damage it can do. Here, that means limiting which tabs and actions the agent can touch.

Our take: this is good news with fine print. The tasks OpenAI is targeting are the ones that eat up hours of people's days: copying data between tabs, filling in forms over and over, comparing quotes. Taking that drudgery off people's plates is a small but real version of the abundance we expect from AI in the long run. In the near term, though, the browser is the new battleground for agents, and Google holds the high ground. OpenAI gains users it could not win with Atlas. It also depends on a competitor's platform. The winner will be whoever earns enough trust for users to let an agent click "send" for them, not whoever has the smartest model. That trust has to come from safeguards that work, not from slick demos.