OpenAI has introduced ChatGPT Atlas, a new browser that integrates ChatGPT directly into the browsing experience. It’s a significant step toward what the company calls an “AI-native web,” where the assistant isn’t a separate tab or tool but part of how you move through the internet itself.

Atlas blends a standard browser interface with ChatGPT’s conversational and task-based features. You can ask a question, enter a URL, or request a task from the same starting point. The model can pull live information from the web, summarize pages, or automate basic online work without you copying links back and forth between tabs.

Crucially, Atlas uses browser memories to make the assistant context-aware. If you enable them, ChatGPT can remember what you’ve recently browsed and resurface that information later, say, summarizing the job listings you looked at last week or continuing research on products you’d compared earlier. These memories are stored privately within your account and can be viewed, archived, or deleted at any time. Visibility controls appear directly in the address bar, letting users decide when ChatGPT can or cannot see a page.

The most ambitious feature is agent mode, which lets ChatGPT act within the browser rather than simply generating text. Users can instruct it to open tabs, click through workflows, or complete repetitive tasks such as researching, shopping, or compiling briefs. For now, it’s available in preview for Plus, Pro, and Business users.

OpenAI has also put limits in place. The agent cannot run code, download files, or access your local computer, and it pauses before acting on sensitive sites like banks or email clients. The company acknowledges that these kinds of agents can be tricked by hidden instructions on web pages, and says it’s building adaptable safeguards to counter that.

Atlas introduces more visible privacy controls than many AI-integrated tools. Browsing history, memories, and agent actions can all be cleared manually. Content viewed in the browser isn’t used to train models unless the user opts in via data controls. Parental controls from the ChatGPT app carry over automatically, with options to disable memories or agent mode entirely.

ChatGPT Atlas is launching first on macOS for Free, Plus, Pro, and Go users. It’s also in beta for Business, and Enterprise and Education users can access it if enabled by their administrators. Versions for Windows, iOS, and Android are on the way.

Future updates will include multi-profile support, developer tools, and new ways for apps built on the ChatGPT SDK to appear inside Atlas. OpenAI also encourages website owners to use ARIA tags to help the agent better understand site structure.

Atlas arrives just a little while after Comet, another ambitious rethinking of the browser. Both share a similar mission: bringing AI directly into the browsing experience instead of treating it as an external chatbot. But their philosophies diverge.

Comet focuses on search-native browsing, turning every new tab into a conversation that blends real-time web results, citations, and context in a clean, question-driven flow. It’s about curiosity and exploration and finding and understanding information faster.

Atlas, by contrast, aims to be action-native. It doesn’t just retrieve information; it performs tasks within your browsing session. Where Comet reimagines how you discover content, Atlas wants to redefine how you interact with it, remembering what you’ve done, automating follow-up steps and even taking actions like purchasing or organizing.

In that sense, Comet feels like an AI-powered navigator, while Atlas positions itself as a capable co-worker sitting in the browser with you.

Atlas is OpenAI’s attempt to fold AI assistance into the everyday mechanics of browsing rather than keeping it as a separate overlay. It’s an early look at how “agentic” web use might feel: part search engine, part task runner, and part memory system. Whether users embrace the trade-off between convenience and visibility will determine if Atlas becomes a niche experiment or the next default way to browse.