OpenAI’s post on X went up on Thursday evening, in the flat register the company uses for product news: “We’re bringing the power of computer use to anyone who uses Chrome.” By then the extension it pointed to had been installed by roughly seven million people, according to its listing in the Chrome Web Store. What those users are being offered is a different kind of software from the sidebar that has sat in their browser for a year.
In Chat, ChatGPT can now look across the tabs a person has open and pull information out of them, comparing listings, articles or documents without being handed the text one page at a time. In Work mode, the agent OpenAI built for longer tasks, it can act on those tabs: send messages, fill in forms, download files, and sort a row of open tabs into groups.
The sidebar carries the full history of a user’s conversations, so a task begun on a phone can be continued while a page is open. The company said the feature is available on every ChatGPT plan, including the free tier, which is unusual for a capability that consumes a lot of inference.
Setup runs through the desktop app. A user installs the extension, enables computer use in settings, and selects the extension inside a Work or Codex conversation. The extension also works in Chromium-based browsers such as Edge, Brave and Opera, and it can read or act on sites where the user is already signed in, among them Gmail, LinkedIn and Salesforce, provided the permissions are granted.
The strategic read is easier to see against the past year. OpenAI launched its own browser, Atlas, in October 2025 and pitched it as a challenge to Chrome.
In March it said it would fold Atlas, the desktop app and its Codex coding tool into a single application. In July it said Atlas would be deprecated, and on Aug. 9 the browser stopped working. Users who wanted help while browsing were pointed to the extension.
That is a concession about distribution. Persuading people to change browsers is a slow, expensive fight against a default that ships with the operating system on many machines and is installed on billions of others. An extension asks for a click instead. The capability OpenAI spent a year building into its own browser now runs inside Google’s, which is where the tabs already are.
Google has been moving in the same direction. Gemini is built into Chrome, and the company has been adding agentic features that can act on pages rather than only summarize them. The two companies are now selling the same promise to the same users in the same window, which is a different competition from the one over model benchmarks.
The capability itself is less novel than the distribution. Computer-use agents that click and type have been demonstrated for two years, and the hard part has never been the demonstration. It is reliability across sites that were not built for a machine reader, and it is security.
Security researchers have warned for as long as these agents have existed that a browser agent is exposed to instructions hidden in page content. A prompt injection buried in a document or a product listing can redirect an agent that is already authenticated to a user’s mail or calendar. The more sessions the agent can reach, the more a single manipulated page is worth to an attacker.
OpenAI’s documentation addresses this through permissions and by requiring the user to enable the capability, which puts the decision on the person who installs it. That is a familiar allocation of risk in consumer software, and it tends to hold until something goes wrong at a scale that regulators notice.
For enterprises, the calculation is different. A tool that can act inside a signed-in Salesforce or Gmail session is useful and also a data-governance question, and IT departments that blocked earlier browser agents are unlikely to unblock this one because it arrived in a Chrome tab.
Analysts who follow the assistant market said the free-tier inclusion is the part with consequences. Charging for agents has kept them in the hands of early adopters; giving the capability to every account makes it a default that ordinary users encounter without seeking it out. The cost of that inference is a bet that usage produces enough conversion and enough data to be worth the compute.
The pricing decision also tells you where OpenAI thinks the constraint sits. Inference for an agent that reads twenty tabs and fills a form is expensive, and giving it away suggests the company expects the cost to fall faster than usage rises, or that it values the behavioral data more than the margin on a subscription.
The comparison that matters may be to the address bar. A generation of users learned to search by typing into a box at the top of the browser and accepting what came back. An assistant that can read the tabs and then complete the form is a different proposition, and it will be adopted or rejected one task at a time, by people deciding whether to let it press the button.
