An AI browser is a web browser — or a browser feature — built so an AI assistant can read a page, click links, fill out forms, and carry out multi-step tasks across the open web on someone’s behalf, rather than just answering questions about text a person pastes in. The person still sets the goal — “compare these three flights,” “find and book a table for four” — and the assistant does the clicking, typing, and reading needed to get there.
What it is
The idea shows up in three different shapes today. The first is a standalone AI browser built from the ground up around an assistant, such as Perplexity’s Comet, which pairs a normal Chromium-based browser with a sidebar that can act on the page. The second is a feature bolted onto an existing browser: Anthropic’s Claude for Chrome and Microsoft’s Copilot Mode in Edge both add an assistant to a browser people already use, rather than asking them to switch. The third is a browser that runs on a remote server rather than the person’s own device, controlled entirely by an AI agent that carries out web tasks in the background and reports back when it’s done — the model OpenAI moved toward after retiring its standalone ChatGPT Atlas browser (see “In the news” below).
All three sit on the same underlying idea: an AI agent that can act on a live web page, not just read a static copy of it.
How it works
An AI browser gives its underlying model a structured view of the page — the text, buttons, and form fields, sometimes alongside a screenshot — much like the view a screen-reader gives a person who can’t see the screen. The model decides what to click, type, or scroll next, the browser carries out that action, and the model reads the result before deciding on the next step. This loop repeats until the task is done or the assistant asks the person to confirm something risky, like a purchase. It’s the same core technique behind computer use, applied specifically to a browser tab — and because it runs inside the person’s own logged-in session, it can act on accounts and saved passwords the way the person themselves could.
Why it matters
The appeal is obvious for repetitive, multi-step web work: comparing prices across tabs, filling out the same form on several sites, or pulling numbers out of a dozen pages into one summary. But the category is young and still consolidating — OpenAI’s decision to fold Atlas into a Chrome extension and a separate cloud agent, only ten months after launching it as a standalone app, shows how unsettled the product shape still is.
The bigger open problem is security. Because the assistant treats whatever is on the page as an instruction to consider, an attacker can hide text on a website — invisible to a person, but readable by the model — that tells the assistant to leak data or take an action the user never asked for. This is called prompt injection, and independent researchers have demonstrated it against several agentic browsers, in some cases finding ways around the same-origin policy that normally stops one website from touching another’s data. Every major vendor treats this as unsolved rather than fixed, which is why most AI browsers ask for explicit confirmation before a sensitive action and are best kept away from logged-in banking or email sessions for now.
Cost varies by product. As of mid-2026, Perplexity’s Comet is free to use on Mac, Windows, iOS, and Android. Claude for Chrome remains a beta feature bundled into Anthropic’s paid Claude plans, starting at $20/month for Pro.
In the news
OpenAI’s own retreat from a standalone AI browser is a case in point: see our report on OpenAI sunsetting ChatGPT Atlas by August 9, 2026.