The AI Browser Power Play
Earlier this month, OpenAI announced ChatGPT Atlas, “a new web browser built with ChatGPT at its core,” which it is promoting to the service’s hundreds of millions of users. Unusual for OpenAI, which has a history of rushing to beat competitors to market, Atlas launched into a crowded marketplace for AI-assisted browsers. Heavyweights Google and Microsoft are already integrating chatbots and other LLM features into Chrome and Edge. Opera has been building AI features into its apps since 2023, while AI search-engine Perplexity released its own browser, Comet, promoting “agentic” features earlier this year. In late 2024, the Browser Company, the maker of Arc, announced plans to shift focus to a new AI-centric browser called Dia. Even Firefox has chatbot integration now, leaving Apple’s Safari, with its few scattered Apple Intelligence features, as the last major browser without a new chat interface growing in or on top of it.
Each of these browsers has a different take on what it means to integrate — or build a browser around — generative AI, but a few common features have emerged. You’ve got easily accessible summarization, editing, and lookup capabilities, of course, as well as different abilities to (borrowing Dia’s words) “chat with your tabs.” In interface terms, they mostly seem to have settled on a right-hand sidebar for chatting, which may end up being this era’s version of the “omnibox,” or tabbed browsing, and which can accommodate the wide range of relationships people have developed with services like ChatGPT, from occasional search replacement or word generator to consultative companion to a layer through which their entire digital world is interpreted. Finally, in architectural terms, we’re mostly talking about browsers built on top of Chromium, the Google-supported open-source project.
If there are major features in Atlas that weren’t already teased or tested in one of these browsers, I haven’t had much luck finding them. In the demo above, even Sam Altman struggled to muster enthusiasm as his employees showed off text summarization, web search, and the automatic chat sidebar. Most of the time, it feels, at least in its current state, like a browser made by OpenAI staffers who were tired of opening a competitor’s product to use their own. In regular usage, it’s less of a new take on browsing than it is a fuller interface for using ChatGPT. That makes sense: OpenAI is attempting to do in 2025 something like what Google did with Chrome’s launch in 2008, guiding users into a relationship with the company deeper than a single product, and much harder to leave. In doing so, it’s also targeting Google more directly than ever. When you open ChatGPT in Chrome (or any other browser), you’ll now get a prompt to try Atlas. When you open Atlas, you’ll see an interface that’s far Googlier than your typical ChatGPT window, with default tabs for Search, Images, Videos, and News. In its launch demo, ChatGPT talked about how popular “search” behaviors are for ChatGPT users, and spent some time showing how the browser can be used to work within Google docs. It wasn’t subtle. They’re coming for Chrome.
This is, again, pretty straightforward tech-firm behavior with plenty of precedent: A company with a major web-based product and ambitious omnidirectional plans wants to own its distribution. This sort of thinking isn’t just how Google ended up with a browser but also with a mobile operating system; it has inspired less successful adventures, too, like Facebook’s failed attempt to make a phone. As a solid browser with a few perks for heavy ChatGPT users, Atlas would be yet another example of OpenAI’s transition from a slippery, speculative nonprofit to a more conventional sort of major consumer and enterprise tech company — from a firm that talks about ASI to a firm that also has lots of KPIs — except for one thing: Agent Mode.
OpenAI’s announcement leaned heavily on Agent Mode, which the company says can help with “researching and analyzing, automating tasks, and planning events or booking appointments while you browse,” and with good reason: It’s pretty wild to watch. Like previous “computer use” features, including OpenAI’s Operator and ChatGPT Agent, Atlas’s Agent Mode effectively takes control of the browser interface, navigating pages, moving a mouse cursor, filling out forms, and narrating the entire process in a sidebar. As a demo, it’s powerful: Look, the computer is using itself! When I asked it to buy me some wool socks in a certain size and color, it navigated to REI’s website and added something pretty close to what I asked for to the cart before asking me to log in; same for a specific type of floss, although it took the liberty of upping the order to four boxes to meet Walmart’s free-shipping threshold. When I asked it to find a flight for this weekend, it spent a few minutes browsing through travel aggregators, describing its strange process in a sidebar in deep detail: “I need to consider the best way to handle the flight search and booking, ensuring I find the most relevant and convenient options for the user. Let’s get this just right,” it printed before creating a Python script to figure out what day it was and calculate what I meant by “this coming Saturday.” It got there in the end, sort of, in that it found a flight that matched my request. Neat! And a task I will not be assigning it for real.
As a performance, computer use is impressive. For many Atlas users, it will also be their first encounter with an agentic tool that can plausibly execute multistep tasks on their behalf. (Programmers will be more familiar with the sensation if they’ve used AI coding assistants, and anyone who has used Perplexity’s Comet or Opera’s Neon will have tried something nearly identical already.) As a feature, it still sits somewhere in AI limbo, where the suggestion that such a tool could one day be useful — if it were faster, smarter, knew more about you, etc. — remains far more powerful than the tool itself.
In other words, like a lot of AI products and companies, Agent Mode benefits from the tailwind of belief in inevitable, continual, exponential improvement. For now, it’s something many users will test, play with, and perhaps incorporate into their general models of the future, if not their actual workflow. (OpenAI currently describes it as a “preview” feature, and it’s only available to paid users.) Browser Company founder Josh Miller, who experimented early on with similar features in Dia, recalled agentic browsing feeling initially dazzling but eventually like a bit of a dead end, or at the very least a strange and cumbersome way to interact with the web and AI. “Training an AI model to click around a computer like a human is akin to putting a print magazine on the early web,” he told me. “It feels futuristic, but it’s unimaginative and not native to the technology.”
As I spent time watching ChatGPT bonk its way through various web interfaces, I also found myself thinking of self-driving cars. A browser that pretends to be a person at the input level — moving a cursor, scrolling a human GUI — felt less like a Waymo, in which an unattended steering wheel turns as a result of actions taken by systems closer to the road, than a regular car with a humanoid robot sitting in the driver’s seat. Again, it’s pretty interesting to watch! But it also makes you wonder: Why are we doing this that way? Aren’t there better ways for machines to talk to each other?
The answer, Miller suggests, comes down to the “incentives of their makers more than the intrinsic value of the technology.” (The Browser Company was recently acquired by software-maker Atlassian.) For OpenAI, building systems that can execute complex commands on behalf of users is the whole ball game — it’s the path to wide-ranging automation and/or AGI, depending on which definition of the term the company is going with that day. An AI model that can provide useful information in a chat window, or handle tasks clearly outlined by the user, is a useful product, but the prospect of an AI model that can productively and proactively interact with the world around it in ways comparable to a human — or, more specifically, an employee — is where trillion-dollar valuations come from.
To get where it wants to go, though, OpenAI has a number of challenges. Some are widely discussed and frustratingly hard to pin down, revolving around benchmarks, varying definitions of model capability, and predictions about scaling. Others are more banal: To answer questions more usefully, for example, chatbots tend to do better if they have more data about users; likewise, to execute tasks on their behalf, they need to operate in an environment where users are logged in to the various services they use to work and live their lives. They need a breathtaking amount of access and permission, in other words. ChatGPT isolated in a chat window doesn’t have that, and it takes a long time to draw out of users, if they’re willing to offer it at all. ChatGPT as a default browser — authenticated in dozens of different sites, payment methods at the ready, or perhaps even logged into a work environment — does. (Such agents also create, as many in the AI space have pointed out, a potential security nightmare.)
That a given service or piece of software could be more powerful with access to all the other software you use is sort of a tech truism — look no further than the thousands of times you’ve been asked by one company for “permission” or “access” to data from another for evidence. But it is usually managed through official partnerships and software APIs. By installing itself at the browser level, OpenAI is looking for the mother of all loopholes: That’s where people who use computers, particularly for work, spend all their time, and through which vast quantities of valuable information flow in and out. Also, if you’re a company hoping to train your models to replicate a bunch of white-collar work, millions of browser sessions would be a pretty valuable source of data.
It really does feel like working with a colleague with a bad attitude when using ChatGPT Agent mode.. having to discuss back and forth to complete relatively mundane and simple tasks is a bit ridiculous, especially considering that the only kind of tasks Agent-mode seems capable… pic.twitter.com/ucZL7ZnkzW
— Oliver Andersen (@ofandersen) October 23, 2025
It would be amazing for OpenAI, in other words, if lots of people found Agent Mode useful and used it at scale, but this is far from a sure thing; likewise, it would depend on a bunch of other companies tolerating the gradual botification of their users without taking the sorts of countermeasures that, say, Ticketmaster does against automated snipers (also, you can’t make money showing ads to a ChatGPT agent!). Instead, I think, Atlas can be understood as a statement of intention, or as a preemptive power play. Before launching its browser, OpenAI announced that it was partnering with a number of companies to build their services directly into ChatGPT, so users could, for example, use the chatbot to create a playlist that would then show up in Spotify, or transition from a chat session directly into a productivity tool like Canva. As a way of working with chatbots, it was far less clunky than the self-driving browser. But it also requires willing partners, and the lineup at launch was fairly thin, consisting largely of companies with nothing to fear from OpenAI, little to lose, or that are already major OpenAI customers.
In OpenAI’s ideal world — much like Google’s, Meta’s, and Apple’s, to be fair — every other tech product exists as a subordinate app inside its own platform, letting ChatGPT users seamlessly navigate their digital lives through the company’s interface, its agentic capabilities unencumbered by security concerns, privacy, competition, or interfaces designed for humans rather than machines. Plenty of companies will be eager to partner with a popular service like ChatGPT. Others — particularly those who see OpenAI, or AI in general, as a potential threat to their businesses or their proprietary data as a protective moat — will be more cautious. OpenAI’s sources of leverage are its large user base and its narrative of inevitability. Agentic browsing is an attempt to fuse the two: A feature that may soon be available to and used by lots of people, that may soon be better than it is.
If building an ecosystem for other companies is a polite invitation into such a future, in other words, Atlas functions as something more like a threat, setting up an easy-way-or-hard-way scenario for potential partners or competitors: If you’d rather not work with OpenAI for now, fine. In the meantime, it will be trying to take over your relationship with your users or customers, deceptively and without your consent. Capisce?