Hayden Field, reporting for The Verge:

OpenAI’s next move in its battle against Google is an AI-powered web browser dubbed ChatGPT Atlas. The company announced it in a livestreamed demo after teasing it earlier on Tuesday with a mysterious video of browser tabs on a white screen.

ChatGPT Atlas is available “globally” on macOS starting today, while access for Windows, iOS, and Android is “coming soon,” per the company. But its “agent mode” is only available to ChatGPT Plus and Pro users for now, said OpenAI CEO Sam Altman. “The way that we hope people will use the internet in the future… the chat experience in a web browser can be a great analog,” Altman said…

[Adam Fry, the product lead for ChatGPT search,] said one of the browser’s best features is memory — making the browser “more personalized and more helpful to you,” as well as an agent mode, meaning that “in Atlas, ChatGPT can now take actions for you… It can help you book reservations or flights or even just edit a document that you’re working on.” Users can see and manage the browser’s “memories” in settings, employees said, as well as open incognito windows.

Atlas is not a novel concept. In the last few years, there have been many browsers that integrate artificial intelligence into the browsing experience:

  • Arc, by The Browser Company, which was recently acquired by Atlassian, the company that makes Jira. Arc gained AI features way before they were popular.
  • Dia, The Browser Company’s replacement for Arc, which more directly mirrors Atlas.
  • Gemini in Chrome, by Google, which aimed to compete with Arc and Dia.
  • Microsoft Copilot in Edge, which seems to be universally hated.
  • Comet, by Perplexity, the search engine hardly anyone uses, yet decided to put in an offer to purchase Chrome higher than its entire net worth.
  • And now, Atlas, by OpenAI.

Atlas is, per an OpenAI engineer, entirely written in SwiftUI for the Mac and uses Chromium, an open-source browser platform owned and made by Google. (Chrome, Dia, Arc, Edge, and Brave use Chromium, just to name a few.) The browsing experience is unremarkable and similar, if not slightly worse, than its competitors because it is the exact same browser. These AI companies are not making new browsers — they’re writing new skins that go on top of the browser. Atlas just ditches Google Search in favor of ChatGPT (set to “Instant” mode) and provides a sidebar to open the assistant on any web page, effectively providing it context. This is both Dia’s and Comet’s entire shtick, and they had their figurative lunches eaten by OpenAI in an afternoon. Dia is even powered by GPT-5, OpenAI’s large language model, and structures its responses similarly to ChatGPT.

I find the experience of using ChatGPT in Atlas, however, to ironically be subpar. Unless a user types in a URL or manually hits the Google Search button in the New Tab window, all queries go to ChatGPT, which answers the question rather slowly. No custom instructions have been provided from OpenAI to prefer searching the web for queries, displaying images or video embeds, or providing brief answers like Google’s AI overviews. It is the normal version of ChatGPT in the browser, and chats even sync to the standard ChatGPT app. At the top are some tabs to expressly show search results piped in from Google, as well as images, videos, and news articles. These results are just one-to-one copies of Google’s, and ChatGPT does no extra work. The search experience in Atlas is terrible and easily worse than Dia or even Google. That’s a shame, because I still find that muscle memory leads me to instinctively use Google whenever I have a question, even though its AI overviews use a considerably worse model than ChatGPT.

The sidebar, which can be toggled at any time by clicking the Ask ChatGPT button in the toolbar, adds the current website to the context of a chat. Highlighting a part of a web page focuses that part in the context window. Aside from the usual summarization and chat features, there’s an updated version of Agent that allows ChatGPT to take control of the browser and interact with elements. Whereas Agent in the ChatGPT app works on a virtual machine owned by OpenAI, this version works in a user’s browser right on their computer. In practice, however, it is useless and often fails to even scroll down a page to read through it. I certainly wouldn’t trust it with any important work.

Atlas is not a good browser. The best browser on macOS today is Safari, and the best Chromium one for compatibility and AI features is Dia, with an honorable mention to Arc for its quirkiness. Anything else is practically a waste of time, and even though I find Atlas’ design tasteful, it’s too much AI clutter that adds nothing of value, especially to this already burgeoning market. And not to mention, the browser is susceptible to prompt injection attacks, so I wouldn’t use the AI features with any sensitive information. I’m sure OpenAI knows this, too, but it decided to release the browser anyway to do some data collection and analyze people’s browsing habits. It’s not a profit center, but a social experiment. The solution is for OpenAI to just make ChatGPT search better1, then offer it as a browser extension to redirect queries from Google, but my hopes aren’t high.


  1. When I mean better, I mean results should follow the structure of Google Search, which has immense staying power for a reason. An overview at the top, some images or visual aids, then 10 blue links for further discovery. That’s a great formula, and OpenAI could make ChatGPT a much better search engine than Google in probably a day’s work. And if it really wanted, it could make that version of ChatGPT Search exclusive to Atlas. ↩︎