ChatGPT Atlas: OpenAI’s 2025 Browser and Its Agent Demos

OpenAI has introduced ChatGPT Atlas, an AI-native browser, and I am impressed by how it brings an assistant into navigation and browser control. I am reacting to the company’s launch demonstration rather than reporting my own extended test. Atlas was later deprecated: OpenAI said it was scheduled to stop working on August 9, 2026. The product and plans described here are the October 2025 launch.

Sam Altman and the people who built Atlas presented several use cases. If you have used Perplexity Comet, some ideas will feel familiar, but I think OpenAI’s interface is especially polished. The home screen resembles a conventional browser, with ChatGPT conversation history on the left.

The main bar accepts either a URL or a prompt, and suggested actions include agent mode, where ChatGPT can operate the browser for you. I see memory as a differentiator: in the launch interface, suggestions could reflect prior activity. Browser memories were optional, and the assistant’s access depended on user settings and permissions.

One example searches browser history when you remember the substance of a page but not its title or address. I have found a similar approach useful in email: describing what a message was about can be easier than remembering its sender or exact keywords.

The demonstrator asks for a document about Atlas core design. ChatGPT searches the history and returns a response pointing to the page, alongside tabs for other kinds of results. The video also includes a sponsored segment from Abacus.AI; its advertised products and price were those presented at the time.

The ChatGPT sidebar stays beside a webpage, allowing questions about what is on it. In the GitHub example, the demonstrator asks for a summary of a task and its acceptance criteria, then asks whether the change is safe to cherry-pick into a release candidate.

The model describes the diff as mostly visual and non-destructive and judges the change low risk. That is a useful example of contextual assistance, but a model’s reading of a diff cannot establish that a release is safe: developers still need appropriate review and testing.

The next search concerns a new Paul Thomas Anderson film. The first view offers a ChatGPT synthesis with links, images and a discussion of reviews. Other tabs provide a more familiar search-results view, including images, videos and news.

I suspect Google powers the conventional results because of a Google link and the speed of the results; that is my inference from the demo, not a confirmed technical explanation. I use ChatGPT for most searches, but a traditional search can be faster and more direct when I already know the destination.

Clicking a result opens the webpage while the ChatGPT response moves into a sidebar, so the conversation can continue alongside the source. The demonstrator opens a film review and asks for a summary in five words or fewer.

Next comes a Gmail draft. The demonstrator selects the text, invokes ChatGPT from inside the page and requests a cleanup. The resulting text can be edited, accepted or cancelled rather than being silently sent.

The more ambitious example asks the agent to work across a Google document and a Linear tracker. It should comment on missing assignments in the document, tag the relevant people by email, and create and assign issues for the current week’s tasks in the tracker, which it finds through browser history.

When the agent takes over, a visible cursor shows its actions. I think that visibility is a deliberate design choice: a user can follow what is happening. The video also contains a sponsorship from Dell Technologies; the demo itself does not establish that every action an agent takes is correct or safe.

The presenter moves on while the first task runs, illustrating the appeal of delegating a multi-step job instead of watching it continuously. In the second example, ChatGPT identifies ingredients for a recipe serving eight, and the user asks it to order only the meat and produce.

The demonstration offers logged-in and logged-out choices before the agent proceeds. Berman contrasts using a local browser with the earlier Operator experience, which ran a separate cloud environment and could require repeated authentication. Access is governed by the user’s choices; browser content and credentials are sensitive, and untrusted pages can attempt to manipulate an agent.

The agent opens a Safeway storefront on Instacart and visibly searches for and adds ingredients such as peppers, green onions, garlic and ground beef. The user can interrupt or take control. In this particular demonstration, the agent prepares a cart and waits for the person to review and complete the purchase.

Returning to the earlier task, the presenters show the agent’s work after about three minutes: issues in Linear and comments in the Google document. This is a launch demonstration, so the examples show intended workflows rather than measuring reliability across all websites.

At launch, Atlas was available on macOS and Berman said Windows would follow. He described agent mode as available to Plus and Pro users; OpenAI’s launch announcement also included Business in the preview. These details are historical: OpenAI later deprecated Atlas. Berman planned to test it further in a subsequent video.