Skip to content
8 min read
Worn white robots with glowing pink eyes, one central robot displaying a pink-tinted icon for ChatGPT Atlas, in a dark alley with pink neon circle

OpenAI’s ChatGPT Atlas Browser Needs Work

Like many people, I tried OpenAI’s ChatGPT Atlas browser last week. I immediately made it my daily driver, seeing if I could make the best of it. Tl;dr: it’s still early days and I don’t believe it’s quite ready for primetime. But let’s back up a bit.

The Era of the AI Browser Is Here

Back in July, I reviewed both Comet from Perplexity and Dia from The Browser Company. It was a glimpse of the future that I wanted. I concluded:

The AI-powered ideas in both Dia and Comet are a step change. But the basics also have to be there, and in my opinion, should be better than what Chrome offers. The interface innovations that made Arc special shouldn’t be sacrificed for AI features. Arc is/was the perfect foundation. Integrate an AI assistant that can be personalized to care about the same things you do so its summaries are relevant. The assistant can be agentic and perform tasks for you in the background while you focus on more important things. In other words, put Arc, Dia, and Comet in a blender and that could be the perfect browser of the future.

There were also open rumors that OpenAI was working on a browser of their own, so the launch of Atlas was inevitable.

Browser War 2.0

For anyone who lived through the early days of the web, there was the so-called Browser War, a fierce competition between Netscape and Microsoft Internet Explorer. A favorite pastime of the tech industry at the time was tracking the dwindling market share of Netscape as Microsoft (illegally) bundled IE with Windows and brute-forced its way to dominance. Google would launch Chrome in 2008 and it steadily climbed the charts until it overtook IE around 2012.

To be fair to Google, Chrome was (and still is) a good product. Unlike Internet Explorer, it was fast and uncluttered. Each tab ran in its own process, which meant if one site crashed or hung, it didn’t crash your entire browser. (Yes, that was a huge issue at the time.)

But in 2025, AI wants to bust out of the confines of its isolated tab and ride along as your sidekick as you go about your business on the internet. Context makes AI a better assistant.

So while Google slept on deeply integrating Gemini into Chrome and making it known, and Microsoft poorly marketed Copilot in Edge, upstarts like Dia and Comet came about. And now, Atlas.

Initial Impressions

The Atlas UI is minimal. Its homepage is essentially ChatGPT but the ask text field is an omnibox that can distinguish between URLs, search keywords, and questions.

Interestingly, when typing a search query like “whole foods,” Atlas will display search results first and answers second. The reverse of what Google search does these days where they’re prioritizing AI overviews over search results.

Dark ChatGPT Atlas page showing search results for "whole foods" (Whole Foods Market, Careers, Wikipedia, Amazon) with a bottom search box.

Search results for “whole foods” yields some traditional results at the top and an answer below them.

One convenience that OpenAI retains is accommodating users who just want to search Google. You can do that by holding the Command or Control key when hitting Enter.

Dark browser window showing a centered ChatGPT Atlas input box with the query "what's a good recipe for pasta bolognese" and search suggestions.

Ask ChatGPT or search Google. Your choice.

The main selling point is that Atlas has context from not only your browsing history as you go, but also your past chats with ChatGPT. If you’ve used ChatGPT since GPT-5 came out, its ability to refer to prior chats and memories is very compelling. I think over time, Atlas having all this context about what you do and what you’re asking about will be very helpful.

The Shopping Cart Test

Taking a page from Comet, Atlas also has an “agent mode” where it can click for you to perform tasks in the browser. I ran it through the same shopping cart test that I performed with Comet last time. While having the Whole Foods page open, I pasted a shopping list into chat and asked Atlas to add the items to my shopping cart. Here’s the prompt:

Add the following to my Whole Foods shopping cart. 

Preferences: 
- Organic when possible 
- Whole Foods house brand when available 

2 carrots 
1 yellow onion 
Greek yogurt, plain, whole milk 
Lactose-free milk, half gallon 
Heavy whipping cream 
Blueberries, pint 
1 lb ground beef (higher fat) 
1 lb ground italian pork 
28 oz can of Cento whole tomatoes 
Small can of tomato paste 
De Cecco spaghetti 
1 bag of broccoli florets

When the agent took over, the page showed an overlay of sparkles—maybe over the top?—and clicked around. I could see its reasoning, including calling out the “node ID” of UI elements on the page.

Amazon search results for "organic carrots" showing three 365 by Whole Foods bagged carrot products, prices, left Rufus chat widget and right cart pane.

Atlas thinking through how to browse the Amazon Whole Foods website.

It took over 11 minutes, but the results were mixed:

  • It added two packages of carrots, not two individual carrots.
  • It added two containers of tomato paste when I only asked for one.
  • Inexplicably, three items were in the regular Amazon cart.

Amazon shopping cart page showing Whole Foods delivery banner, cart items (spaghetti, tomatoes), subtotal $45.09 and right-hand chat listing added items.

After 11 minutes, the agent finished with very mixed results.

For comparison, I ran the exact same exercise on Comet again and it finished in 4 minutes, with two mistakes:

  • It added two packages of carrots, not two individual carrots.
  • It added the wrong brand of spaghetti, choosing the house brand instead of De Cecco as specified.

Why It’s Not Quite Baked

Increasingly, we live in our browsers. Web apps have by and large replaced apps on our computers. Email, chat, spreadsheets, design tools, and more are in the browser. And it makes sense for an AI assistant to be there watching and helping.

But the browser needs to be a good one. By stripping back some of the UI it thought unnecessary, OpenAI hamstrings it. For instance, I can’t get the 1Password extension to work correctly. It’s always the first extension I install because all my passwords are in that vault, which I need to log into my apps. Other niceties that are missing include tab groups and profiles.

Without the basics and without being better than Chrome, it’s hard for me to make Atlas my default browser. I’ve been using Comet as my daily driver since July and while I really miss some of Arc’s features like the sidebar tabs, at least Comet has all the other core Chrome features. Dia has fallen behind in my opinion, but maybe the injection of Atlassian money will help them accelerate.

The Privacy Tradeoff

Tech blogger Anil Dash called Atlas “anti-web” in a scathing review, arguing that search results don’t properly link to websites. He’s worried about plagiarism and attribution.

When I tested it, links appeared at the top of results pages. Maybe the version he used didn’t have them, or maybe he missed them. Either way, his broader concern—that AI assistants create comprehensive surveillance profiles as they follow us around—is worth taking seriously.

Dark-mode web profile showing a music artist overview with a portrait, concert thumbnails, a heading, and bullet-point biography text.

Search results for “Taylor Swift” in Atlas includes links to her official website, Instagram, and Wikipedia page at the top. Then includes her bio.

The privacy tradeoff is real. AI can only be as useful as the context you provide it. For it to become a truly helpful assistant, it needs to know what you do on your computer. There’s a reason that great executive assistants who know the details of your personal and work life are more highly-paid and effective than “virtual” assistants who can only perform simple tasks. Leadership coach Michael Hyatt on what makes a rockstar EA:

A rockstar EA is like a second brain. She knows what you like and don’t like. She knows where you are and where you need to go. She knows when to schedule meetings and when not to. A rockstar EA will gather as much of this information as possible as early as possible—and proactively keep learning.

For LLM-based AI assistants, these privacy tradeoffs are going to be for each of us to decide.

The Future Is Still Being Built

Computers are supposed to work for us, to help us get things done. Too often, we’re working for them. Complicated UIs are partially to blame. But the real culprit? The way software is siloed. Every piece of software that we use, be they web apps, SaaS products, or desktop apps, is purposely walled off from one another. Yes, you can share files or open the same file in different apps, but they don’t actively talk to each other. Quite literally, the left hand doesn’t know what the right hand is doing. Apple tried to solve this back in the ’90s with OpenDoc, a document-centered framework where you’d build documents from modular parts rather than monolithic applications, but Steve Jobs famously put a bullet through its head in 1997.

OpenAI and others are trying desperately to close that gap and break down the silos. When ChatGPT gains knowledge and memories across your tabs and browsing history, it can be more assistive. But of course, not everything we do is web-based. OpenAI announced last week that it’s acquired the company behind Sky, essentially an AI copilot for the Mac. Folding that technology into ChatGPT will break down even more silos.

For now, I don’t think the ChatGPT Atlas web browser is there. It’s still very much a beta in my view. Browsers are too important in our work for them to be super stripped down. Hopefully, OpenAI will make the browser part of Atlas much better. Until then, I still don’t have the Arc-Dia-Comet browser smoothie I’ve been wanting.

Subscribe for updates

Get design insights in your inbox. Sent weekly (or so).