On October 17, 2026, OpenAI shipped a full update to the ChatGPT extension for Chrome. According to The Decoder's report the same day, ChatGPT is now a "full Chrome browser assistant": you call it from the Chrome sidebar without switching to a separate app or the web client; it can read and operate the tabs you have open, and cross-tab tasks such as filling a form across sites or comparing prices between pages run on the Computer Use capability inside Work mode. Every paid ChatGPT tier gets this, not just the top one.
One more piece of the update is easier to scroll past. The Decoder reports that OpenAI's James Sun disclosed the technical details: the extension can see the browser's full context, including content you have highlighted; it can control tabs; and it can access the local file system and installed extensions. Layer on the other official statement, relayed by The Decoder and AIBase: OpenAI will gradually discontinue the standalone Atlas browser and share migration guidance toward the ChatGPT extension. The shape of the whole story changes. OpenAI is no longer trying to build its own browser. It is moving into the one somebody else built and roughly sixty percent of the world already runs.
A note on scope: on October 2 this site covered the open-source Dots browser agent and a resident-agent comparison. Those pieces were about open alternatives and third-party agents side by side. This one is about OpenAI's own product move. They do not overlap.
1. What Actually Changed
Read The Decoder's report against the official documentation at learn.chatgpt.com and the update splits into three layers.
The first layer is location. ChatGPT stops being an app you visit and becomes an assistant that lives in the browser. It sits in the Chrome sidebar, opens with Cmd+Shift+. (period) on macOS, or by right-clicking a page and choosing Ask ChatGPT. You can pull the current tab, or a piece of selected text, straight into the conversation as context.
The second layer is range of action. The official list: research markets, compare sources, extract information from websites; run multi-step operations across pages; fill in forms; open and refine Google Workspace and Microsoft 365 files inside the app; and work inside sites where you are already logged in, with LinkedIn, Salesforce, Gmail and internal tools named as examples. Cross-tab actions of this kind run on Computer Use inside Work mode, available on all paid tiers.
The third layer is the sensitive one, the set James Sun disclosed: full browser context including your highlights, tab control, and access to the local file system and installed extensions. Session threads opened in the browser sync into the desktop app, and either side can reference the other's tasks. Something ChatGPT was halfway through in the browser picks up seamlessly back on the desktop.
Put the three layers together and the conclusion is one sentence: ChatGPT's operating range has grown from a chat box to the whole browser, and then past it.
2. Five Months, Four Steps
This upgrade did not come from nowhere. On the timeline assembled by aicatchup, every step points the same way.
| Date | Event | Source |
|---|---|---|
| 2026-05-07 | Codex Chrome extension launches, Chrome only, some regions excluded | aicatchup |
| 2026-06-16 | Extension opens to the EEA, the UK and Switzerland | aicatchup |
| 2026-07-09 | Codex is folded into the ChatGPT desktop app | aicatchup |
| 2026-10-17 | Upgraded to the ChatGPT browser extension: cross-tab, local files, five browsers | aicatchup |
Five months ago it was a coding extension called Codex, Chrome only. In early July it was absorbed into the desktop app, demoted from a standalone feature to one entry point among several. By mid-October it reappears as the ChatGPT browser extension, with a capability list that has grown from "drive web pages to write code" to cross-tab tasks, local files and five browsers.
The direction of that line is unmistakable: plant a nail in the browser first, then grow the nail into a structure. If you installed the Codex extension in May, this wave is not a new product so much as a doubling of capability in the same spot, which also means the permissions you approved back then are worth rereading today.
3. Three Entry Points, One Capability List
The official documentation defines three entry points: @Computer is the desktop app itself, @Browser is the browser built into the desktop app, and @Chrome is the browser you actually log into every day. All three share the same ChatGPT; what differs is the range they can act on.
The capability list, per the official text: research markets, compare sources, extract information from sites; open and refine Google Workspace and Microsoft 365 files in the app; run multi-step operations across pages; fill forms; and work inside logged-in sites such as LinkedIn, Salesforce, Gmail and internal tools.
A few details are worth singling out.
Browser tasks run inside Chrome tab groups, and the pages belonging to one task are grouped together. That means you can see at a glance which pages ChatGPT currently has open on your behalf, and it means its actions are visible in the browser UI rather than hidden in a black box.
YouTube gets dedicated support: videos are explained and summarized using timestamped subtitle transcripts. When a report needs a quote from a video, this removes the scrub-for-the-line step.
Appshot covers the desktop side: double-press Command, and the frontmost Mac window is attached to the conversation as a screenshot plus readable text. Questions outside the browser come in through this door.
Invocation happens inside a ChatGPT Work or Codex conversation with an @ mention: @Chrome, @Edge, @Brave Browser, @Opera, @Vivaldi. The official example reads: "@Chrome open Salesforce and update the account from these call notes." If the browser is closed, ChatGPT can open Chrome itself.
4. Local Files and Login States: The Two New Lines
Of everything in this update, two lines had not been crossed before, and they deserve to be stated precisely.
The first is the local file system. James Sun's disclosure that the extension "can access the local file system and installed extensions" carries a premise that must be spelled out: it holds only if you have authorized Computer Use in the desktop app. This is not the extension quietly reading your disk. The authorization happens in the desktop application, and the permission is inspectable and revocable at the OS level. On macOS, Computer Use itself requires both Screen Recording and Accessibility permissions; missing either one, nothing runs. The accurate description: grant the permission and it can touch local files; withhold it and it cannot.
The second is login state. The official documentation lists logged-in site tasks as a headline capability, with LinkedIn, Salesforce, Gmail and internal tools as named examples. The weight of this comes from where the extension runs: inside your everyday, already-authenticated browser, it inherits your existing logins. SSO intranets, CRMs, corporate mailboxes, the places ordinary agents cannot reach, it can enter and work. That is also the essential difference between the @Chrome entry and @Browser: the former carries your identity, the latter does not.
Together, these two lines are the real distance between this extension and a chat box bolted onto a web page. But they are also where the risk concentrates. An assistant that can touch local files and sign into your corporate mailbox does not fail cheaply. The official documentation says so itself: page content, selected text and video transcripts are all to be treated as untrusted context, and you should review the page ChatGPT is about to act on before you let it act.
5. Permissions: Ask Before Acting
Capability is half the story; the other half is permission, because this extension's default posture is ask-first.
On a new site, the first action stops and asks. The prompt is based on the hostname, such as example.com, and you face a four-way choice: Allow once, Allow for this site, Allow for all sites, or Decline. Nowhere does the official documentation describe acting without asking.
Two of the four options deserve a second look. "Allow for all sites" and "browser history access" are flagged by OpenAI's own documentation as high-risk choices that expand the permission surface. When the vendor warns against its own convenience option, that warning is worth taking seriously.
To cut down on repeated prompts, there is an allowlist and a blocklist: Settings, then Computer Use, then Manage next to the browser. Domains you no longer want asked about go into the allowlist; domains that must never be touched go into the blocklist. One detail matters here: all five supported browsers share the same site permissions. A site you approved in Chrome is approved in Edge too.
Installation is itself a permission walk: ChatGPT desktop app, then Settings, then Computer Use, pick the browser, Install, approve in the extension store, return to Computer Use and confirm it shows Manage. Every step is an explicit confirmation, with no one-click shortcut through all of them.
The intent behind the design is not hard to read: capability extends past the browser, and the decision stays inside each prompt. The cost is a lot of clicking in the early days, which is exactly the ask-first gate doing its job. Approve all sites once because the prompts annoy you, and you have dismantled the gate yourself.
6. Five Browsers, and How to Pick an Entry
Support covers five browsers: Chrome, Edge, Brave, Opera and Vivaldi, all Chromium-based. Firefox and Safari are not supported, and the official documentation shows no sign of that changing. Opera has one difference: no sidebar chat, though desktop workflows still work.
On choosing an entry point, the official guidance comes down to three sentences. Tasks that need your existing login state, such as SSO intranets, CRMs or corporate email, go through the @Chrome extension. Localhost, preview pages, or anything where a separate setup does not matter, go to @Browser. Services with a dedicated plugin, use the plugin first.
Translate those three sentences and you get one rule: follow the identity. If the work depends on being you, use the browser you actually log into. If it does not, the sandboxed built-in browser is the safer container, if only because it stays separated from your everyday browsing.
Firefox and Safari users have no official answer for now. The practical path is to keep a Chromium-based browser around for agent tasks and keep it separate from the daily driver. Given the shared-permission mechanism described above, that separation is actually the cleaner arrangement.
7. The Strategy Read: Parasite Beats Landlord
Finally, the strategy. First, set the record straight on sourcing: what follows is this article's analysis, not an OpenAI statement. The official wording, as relayed by The Decoder and AIBase, is that OpenAI will gradually discontinue the standalone Atlas browser and share migration guidance toward the ChatGPT extension.
Atlas was OpenAI's own AI browser, previously advanced as a standalone product. AIBase frames the pivot this way: from a standalone browser to a parasitic strategy, OpenAI chose the pragmatic route. Rather than fight for browser share from zero, it moves into Chrome, which already holds roughly sixty percent worldwide.
Resist the temptation to call Atlas a failure. The official position has always been "gradually discontinue and migrate," with no verdict on success; the two confirmable facts are that the standalone-browser road stops advancing, and the Chrome-extension road gets the investment.
Why is parasitism the better play? The browser is one of the highest-switching-cost pieces of software in the world. Users will not move house for a new AI browser, but they will let their current browser take on an assistant. Building your own means winning twice, first the install and then the habit; parasitizing means winning once. And this update's feature list is exactly the full set of furniture needed to live in someone else's house: login state, local files, cross-tab work, five browsers covered.
There is a quieter signal in the timeline too. Codex, an extension named for writing code, grew into a full browser assistant. The contest over AI entry points is sinking from the chat window down to the browser layer. Whoever operates your browser stands between you and the entire internet. With this move, that competition is only beginning.
Closing
Assembled in one place: in capability, ChatGPT has gone from a chat box to full browser context and out to local files; in coverage, from Chrome alone to five Chromium browsers; in posture, from building its own browser back to moving into a home someone else already owns, one that holds sixty percent of the market.
For the ordinary user, the practical advice is three lines. Install through the full approval flow and skip no steps. Day to day, use Allow once and per-site approvals, and save "allow for all sites" for the cases that truly need it. Treat page content as untrusted context, and look at what ChatGPT is about to operate on before you let it act.
As for Atlas, a browser once advanced as a standalone product has ended as a line of migration guidance. Not a surprising ending, but it leaves a question on the desk of every company building an AI assistant: on the browser, do you build one from zero, or move into the one that exists? OpenAI answered with a five-month timeline and an update on October 17.
Related reading
- Five always-on agents, compared: who takes the night shift: Five always-on agents, compared: who takes the night shift
- Open-source Dots: 2,400 stars in three days: Open-source Dots: the browser, handed to an agent
- Google's New Hire: An AI Coworker With Its Own Inbox: Google's new hire: an AI coworker with its own inbox