OpenAI’s latest ChatGPT Chrome extension can pull information from several open tabs and, when connected to desktop Work or Codex, act on sites where the user is already signed in. That could turn a cluster of shopping pages or a form into one delegated task. It also gives ChatGPT a path into accounts already open in the browser.

The current build is dated October 15, but the listing does not identify which features arrived in that version. OpenAI documented multi-tab control and tab mentions months earlier. What is clear now is the combination: tab context in a conversation and browser control for signed-in workflows. These are documented capabilities, not independently tested evidence of how reliably ChatGPT completes everyday tasks.

The most consequential choice appears after installation. By default, ChatGPT asks before using each new website. A user can approve one visit, save permission for a site or choose “Allow for all sites,” which removes future website-access prompts. Website approval is still not a promise that ChatGPT will stop for a final confirmation before every submission or purchase.

What each permission allows

Installing the extension grants browser-level capabilities, potentially including reading and changing website data; ChatGPT separately asks which websites it may use. Browser history requires request-specific permission with no always-allow option, and the store says uploads and downloads require permission too. Desktop app control has its own approval, while page material brought into a conversation falls under ChatGPT’s data controls. None of these grants establishes that every consequential action will receive a final confirmation.

To ask about a page, users open ChatGPT from Chrome’s toolbar or Extensions menu. The side panel can use context from open tabs, and users can mention particular tabs or highlight a passage instead of copying it into ChatGPT. Comparing details across several product pages is an obvious use.

OpenAI’s detailed browser-control workflow starts in the desktop app. Users connect the extension through Settings > Computer Use, start a Work or Codex chat, then select the browser with an @-mention. OpenAI lists data entry and document editing among the supported uses. A form submission or document edit can therefore change the same account the user would otherwise operate by hand.

The browser guide also covers Edge, Brave, Vivaldi and Opera. All five support desktop tab mentions and browser control; Opera lacks the side panel. The Chrome store also mentions ChatGPT on the web, but OpenAI’s detailed control instructions focus on desktop Work and Codex.

  • Changed: The current ChatGPT extension build is dated October 15, 2026. Its listing describes tab context and browser actions, but does not isolate what this build introduced.
  • Added: OpenAI’s May 7 changelog introduced Codex for Chrome with parallel background work across tabs.
  • Added: The July 30 digest documented open-tab mentions and side-chat selections; the August 25 changelog added four more browsers with tab mentions and control.

OpenAI’s pricing page lists Work and Codex with Free, Go, Plus, Pro, Business, Edu and Enterprise plans. But it marks Chrome control and Computer Use as region-limited, while the browser guide adds rollout and workspace restrictions. The documents do not establish identical extension capabilities for every account.

Permission to enter a site is not a guaranteed checkout pause

Once a site is allowed, the next boundary is less definite. OpenAI’s Computer Use guide says ChatGPT may ask before sensitive or disruptive actions. It does not promise a final prompt before every message, form submission, appointment booking, purchase or record change. Individual actions may still be confirmed, blocked or restricted; the documentation simply does not state a universal rule.

That matters because sites may treat clicks, submissions and other signed-in actions as coming from the user’s account. OpenAI warns that pages can contain malicious or misleading instructions and advises users to stay present for payment, credential and account-setting tasks. Users can stop a task or take over at any time.

Even a read-only request can have a privacy cost. OpenAI says it does not keep a separate complete record of extension actions, but page text, screenshots, tool calls and other browser material can be stored when they become part of the conversation. That is not a promise of zero logging. Content from personal accounts may be used to improve models depending on settings; Business, Enterprise and Edu content is excluded from training by default.

The useful part of this extension is also the reason to start cautiously: it can work with the browser session that already holds the information and permissions a task needs. A bounded first job, with one-time site access and the user present for consequential steps, fits the evidence better than assuming a last-minute prompt will catch every costly mistake.