Comet and Atlas: why the AI browser bet failed
Comet and Atlas: why the AI browser bet failed
This is the first of three pieces on AI browsers. The agents that kept the browser is next. Are AI browsers Comet and Atlas already dead? is the last.
For about a year, the serious product bet in consumer AI was a new browser. Chat was a window you pasted into. A browser was the place the work already lived: the tab, the login, the form, the receipt. If the model could see that page and act on it, the address bar would replace the chat box.
Perplexity shipped Comet in July 2025, first on the Max plan, then free worldwide on 2 October 2025. OpenAI shipped ChatGPT Atlas on 21 October 2025, macOS only. The Browser Company's Dia sat in the same conversation. Three products, one sentence: stop copying the page into a chatbot.
That sentence was right about the problem. It was early about the container.
Why a browser felt like the product
Search and chat had split the job badly. You asked a question, got an answer, then opened six tabs to check it. An AI browser promised to collapse that loop. Type the task. The browser searches, opens the pages, summarizes them in place, and, when you allow it, clicks through.
Comet's pitch was research with an agent beside the page. Summaries sat on the article. Agent mode could run a multi-step task. Atlas's pitch was action. OpenAI described a browser that could complete work across sites you were already signed into, with ChatGPT as the sidebar that could take the wheel.
The category had a clean story for investors and for users who were tired of tab hoarding:
- The page is the context. You should not have to paste it.
- The click is the action. You should not have to do the last ten steps yourself.
- The session is the memory. Logins, carts, and dashboards are already there.
By early 2026 the usage stories matched the pitch, at least in third-party trackers. Presenc AI put Comet near 18 million monthly active users in Q1 2026, up from about 2.4 million a year earlier, with agent-task completion climbing as the tool-use stack got less fragile. Perplexity has not published that figure itself. The direction was hard to miss: a lot of people tried a browser whose main feature was a model.
HUMAN Security's May 2026 read of agentic web traffic put Comet at 47% of observed agent traffic and Atlas at 20.3%. Together they were most of the category. Media, commerce, and travel were almost the whole destination list. Product and search routes were about three quarters of the activity. Checkout was a sliver. People were browsing with an agent. They were rarely letting it buy.
Atlas was a browser you had to switch to
Atlas launched on the Mac and stayed there. No Windows build, no iOS, no Android. Bookmarks, extensions, and a decade of Chrome muscle memory do not move because a sidebar is clever. OpenAI later said the plain part out loud: getting people to switch browsers was harder than meeting them inside ChatGPT, which they already open.
The product still taught the company something. A browser session is a better workplace than a chat transcript. Multiple tabs, downloads, logins, and a page that stays put are what an agent needs when the task is "finish this," and a single message box is a poor place to keep that state. Atlas was the prototype of that workplace. It was also a second app to install, patch, and explain.
Comet had a wider door. It went free, it reached phones, and head-to-head tests through the spring often rated its agent mode quicker than Atlas. The iOS app hit the US App Store top tier within two days of the 18 March 2026 launch. Distribution was the advantage. A Chromium fork is still a browser you have to choose every morning.
The bet had a cost the demo hid
An agent that clicks inside your accounts is a different product from an agent that answers a question. In March 2026 a federal court injunction blocked Comet's agents from Amazon's signed-in account pages. The fight was about authenticated sessions, and it spread. Site owners started treating "a browser with a model in it" as a bot with a screen, and blocking rates on agent traffic climbed through the spring even while the two AI browsers still led the category.
There is a second cost, and it sits on the vendor's bill. Driving a real page, with screenshots, tool calls, and retries, is expensive at millions of users. A sidebar summary is cheap enough to bundle. "Control the browser and finish the task" is a metered computer. Comet could grow on the first. The second had to be priced like compute.
That is the rise, stated cleanly. The AI browser won the argument about context. The page should be in the prompt. It lost the argument about distribution. Chrome was already open. ChatGPT, Claude, and Grok were already installed. A third app that was also a browser had to be better than both, every day, for a person who did not want a new default.
What the year actually proved
The useful idea survived the apps. You can see it in what shipped next, which the next piece covers: OpenClaw, Hermes, Grok Bot, Claude Desktop, and the ChatGPT desktop app, each with a browser the agent drives and, in several of them, computer use that clicks the desktop itself. Google's version of the smaller idea, Ask Gemini on the current Chrome tab, is the third piece.
For code review the same split matters. The valuable part of an AI browser was never the new chrome around the window. It was reading the artifact in front of you and acting on that artifact. A pull request is that artifact. ThinkReview sits on the merge request in GitHub, GitLab, Azure DevOps, and Bitbucket, and the model is whichever one you pick in Model selection. The review stays on the page where the change already lives.
Comet's early distribution is Perplexity's own launch record; the Q1 2026 user estimate is Presenc AI's, not an official Perplexity figure. Agentic traffic shares are from HUMAN Security's May 2026 report, as covered by PPC Land. Atlas's launch and later shutdown are in OpenAI's own notice, covered in the next post. Install ThinkReview.