What Is the ChatGPT Browser in 2026?

If you have searched for the ChatGPT browser recently and come away confused, there is a good reason. The thing most articles describe no longer exists. ChatGPT Atlas, OpenAI's standalone web browser, stopped working on 9 August 2026, roughly nine and a half months after it launched.

That leaves a lot of people searching for a product that has been retired, reading guides that were accurate last spring, and reasonably wondering whether ChatGPT can browse the web at all any more. It can, and in some respects it does more than Atlas did. It just does not do it as a browser you download and set as your default.

This guide covers what the ChatGPT browser actually is as of September 2026, what happened to Atlas and why, how to get browser-style behaviour out of ChatGPT today, and the part that matters most if you run a website: what agentic browsing is doing to the traffic that used to arrive from search.

So what is the ChatGPT browser now?

There is no longer a standalone ChatGPT browser. OpenAI retired ChatGPT Atlas on 9 August 2026 and moved its browsing capabilities into two places instead: the ChatGPT desktop app, which now has a built-in browser that can log into accounts, download files and work through sites on your behalf, and the ChatGPT Chrome extension, which reads the page you are already on so you can ask questions about it without leaving your own browser.

OpenAI's own announcement page for Atlas now carries a one-line notice confirming it has been deprecated. The capabilities were not cancelled, they were relocated, and there is also a cloud-hosted browser that agents use to complete tasks without occupying your screen at all.

The practical upshot for most people is simple. If you want ChatGPT to help you with a page you are looking at, install the extension. If you want ChatGPT to go and do something involving several sites and a login, use the desktop app.

It is worth being precise about the terminology, because search results are full of stale guides. Atlas was the browser. ChatGPT search, which answers questions with web citations inside a normal chat, is a different thing and never went away. When someone says the ChatGPT browser now, they almost always mean one of the two surfaces above rather than a downloadable application.

Wondering whether AI assistants are citing your site or your competitor's? Get a free SEO audit and find out where you actually stand.

What happened to ChatGPT Atlas?

The Atlas story is short enough to tell properly, and it is worth understanding because it explains where OpenAI thinks this is all going.

What Atlas actually was

Atlas launched on 21 October 2025 as a macOS-only browser built around ChatGPT. It put a persistent assistant in a sidebar so you could summarise a page, compare products across tabs or interrogate a document without copying anything into a chat window.

Its headline feature was agent mode, available to paying subscribers, which let ChatGPT operate a site autonomously. It also had browser memories, which retained facts from pages you visited so the assistant could refer back to them later, with OpenAI stating those memories were deleted within seven days of being summarised.

By January 2026 it had added tab groups and an Auto search mode that decided per query whether to answer directly or hand you conventional Google results. Windows, iOS and Android versions were announced but never shipped.

Why OpenAI shut it down

OpenAI announced the shutdown on 9 July 2026, giving users about thirty days' notice before Atlas stopped working on 9 August. The reasoning reported at the time amounted to a strategic narrowing: the browser was treated as a feature rather than a destination, and competing with Chrome as a product was judged a distraction from putting the same capability everywhere people already work.

Read uncharitably, a macOS-only browser at a fraction of a percent of market share was never going to displace Chrome. Read charitably, the interesting part was never the chrome around the page, it was the agent, and an agent does not need its own browser to be useful.

There was also a persistent criticism that a browser answering questions instead of showing pages was working against the web it depended on. That argument did not cause the shutdown, but it does frame the tension every product in this category still has to manage.

What users had to do

Anyone with data in Atlas was advised to export bookmarks as an HTML file, save important URLs and open tabs, back up browsing history, and export cookies where possible while treating them as sensitive. ChatGPT conversation history was unaffected, since that lived in your account rather than in the browser.

If you are reading this after the fact and never exported anything, that data is gone. It is worth checking whether an old Atlas install still holds a local profile before you wipe the machine.

How do you get browser behaviour from ChatGPT today?

Work through whichever of these matches what you are trying to do.

1. For questions about a page you are already reading, install the ChatGPT Chrome extension from the Chrome Web Store and sign in with your existing account.
2. Open the page, then open the extension sidebar. It has access to the page context, so you can ask it to summarise, pull out the numbers, or explain a section without pasting anything.
3. For tasks spanning several sites or needing a login, open the ChatGPT desktop app instead and use its built-in browser.
4. Describe the whole task rather than narrating clicks. The agent handles multi-step work better when it knows the goal, so say what outcome you want and what constraints matter.
5. Stay at the keyboard for anything involving a payment, a booking or a submitted form. Agentic browsing is good enough to be useful and not good enough to be unsupervised.
6. For long-running work you do not want to watch, hand it to the agent and let it run in the cloud browser, then check the result rather than the process.
7. If you only want conventional web citations in an answer, plain ChatGPT search on the web or mobile app is still the fastest route and needs no extra install.

None of this requires changing your default browser, which is arguably the point of the reorganisation.

One setup detail worth getting right early: check what the extension can see before you rely on it. Extensions read the active tab, which means anything on your screen is in scope, including pages you would not choose to share. If you work with client data or internal systems, decide deliberately which sites it is allowed on rather than granting access broadly and forgetting.

Not sure which of your pages AI assistants actually read? Request a free SEO audit and get a prioritised list of what to fix first.

Which surface should you use when?

The three surfaces overlap enough to be confusing and are genuinely different in ways that matter.

The Chrome extension suits reading and research. You are in control, you see every page, and the assistant is commentary on top of what you are doing. It is the lowest-risk option and the one to default to.

The desktop app's built-in browser suits work you want done rather than watched, particularly anything behind a login or involving downloads. Because it can authenticate as you, the blast radius of a misunderstanding is larger, which is exactly why it is worth being specific in your instructions.

The cloud browser suits genuinely long tasks where waiting is the cost. It is also the one where you are furthest from what is happening, so it fits well-specified, low-stakes, repetitive work and fits poorly anything requiring judgement.

The honest answer for most people is that the extension covers eighty per cent of what Atlas was used for, and the desktop app covers the rest.

If you were an Atlas user specifically for agent mode, the desktop app is the closest replacement and the transition is mostly a matter of relearning where things live. If you used it mainly for the sidebar, the extension is a straight swap and arguably an improvement, since you keep your own browser, your own extensions and your own profile.

What are the limitations?

Several, and they are the same limitations that applied to Atlas, which is part of why the standalone browser was hard to justify.

Prompt and memory injection remain unsolved. A page can contain instructions aimed at the assistant rather than at you, and an agent reading that page may follow them. Atlas had documented vulnerabilities of exactly this kind, and moving the capability into a different container does not fix the underlying problem. Anything with access to your logged-in sessions deserves scepticism.

Agentic browsing is slow compared to doing it yourself. For a task you have done before and know the path through, you will beat the agent comfortably. The value shows up in unfamiliar or tedious work, not in things you are already fast at.

It is bad at ambiguity. An agent that encounters an unexpected consent dialog, a changed layout or a genuine judgement call will usually either stop or pick something. Neither is what you wanted.

It is also quietly expensive in attention. Supervising an agent is not free, and a task that needs checking at four points is often worse than doing it yourself, because you paid the cost of the work and the cost of the review. The tasks worth delegating are the ones where you genuinely do not need to watch.

There is also a real dependency question. Atlas existed for under ten months. Building a workflow around any single AI surface is a bet on that surface still existing next year, and the Atlas shutdown is a reasonably clear data point about how quickly that can change.

Finally, from a publisher's perspective, these tools read your site without ever showing a visitor to it. That is not a limitation for the user; it is a limitation for you, and it is the next section.

Watching organic traffic slide while AI answers take the clicks? Start with a free SEO audit and see which pages are still earning visits.

What does this mean for your search traffic?

This is the part that matters if you own a website, and it does not change with the Atlas shutdown. If anything it accelerates.

When an assistant reads three pages and synthesises an answer, your page was used and your site was not visited. You may still get credit in the form of a citation, and citations do convert, but the ratio of reads to sessions is moving in one direction. Treating your analytics as a complete picture of how often your content is consumed is now actively misleading.

What that argues for is straightforward and unglamorous. Be the source that is easy to extract from: answer the question near the top, keep claims specific and attributable, use headings that match how people actually ask, and keep facts current enough that an assistant citing you is not repeating something stale. Articles about products that no longer exist are precisely how a site loses trust with both readers and models.

This page is a small example of the dynamic. A large volume of people search for the ChatGPT browser every month, and most of the content answering them describes a product that was switched off in August. Being the page that says so plainly is worth more than being the page that ranks for the old version of the question.

There is a maintenance implication in that. Content covering fast-moving products needs a review cadence, not just a publish date, because the cost of being wrong is no longer a mildly dated paragraph. It is an assistant confidently repeating your stale claim to someone who will never visit to check.

It also argues for caring about brand-level mentions rather than only rankings. Whether an assistant names you when someone asks for a recommendation in your category is becoming as commercially relevant as where you sit in a list of ten blue links.

How does it compare to the alternatives?

Atlas is gone but AI browsing is not, and the field is now more interesting than it was.

Perplexity's Comet browser is the closest thing to what Atlas was trying to be, and it survived the year Atlas did not. Comet and its assistant are available across Perplexity's tiers including the free one, with the more autonomous agent capabilities gated to paid plans. If you specifically want a browser built around an AI assistant rather than an assistant bolted to a browser, this is the obvious place to look.

Google has been folding Gemini into Chrome directly, which is the strategically strongest position in the category simply because Chrome is already where most people are. The counter-argument is that a feature inside a browser owned by a search company will always be shaped by that company's interest in keeping search intact.

Microsoft's Copilot in Edge follows the same logic for Windows and Microsoft 365 users, and is the path of least resistance if your work already lives in that ecosystem.

A handful of smaller AI-first browsers also continue to exist, generally aimed at developers or at people who want more control over which model handles a request. They are worth a look if you have a specific reason, and worth skipping if you do not, because the switching cost of a browser is mostly paid in small daily friction rather than up front.

The pattern worth noticing is that the two standalone AI browsers were built by companies without a browser to defend, and one of them has already closed. Capability is consolidating into browsers people already use, which is precisely the conclusion OpenAI reached.

Final Thoughts

The short version: the ChatGPT browser is no longer a browser. Atlas was a nine-month experiment that ended in August 2026, and its useful parts now live in the ChatGPT desktop app, the Chrome extension and a cloud browser your agents can drive. For most people the extension is the answer, and it takes two minutes to set up.

The larger shift is worth sitting with. Whether the assistant lives in its own browser or in yours is an implementation detail. What matters is that an increasing share of the web is now read by software on a person's behalf, which changes what it means to be findable.

The Atlas shutdown is also a useful reminder to hold any single tool loosely. Nine months is not long, and the sensible posture is to get value from whatever surface works today without rebuilding your process around it.

If you want to work on that side of it, our step-by-step checklist for ranking in ChatGPT, Claude and Perplexity covers the practical work of becoming the source these assistants cite, which is the durable version of this opportunity regardless of which browser wins.

Ready to find out whether AI search is finding you? Get your free SEO audit and see exactly what is holding you back.