What happened to ChatGPT Atlas, and what replaced it?
OpenAI deprecated the Atlas browser and it stopped working on 9 August 2026. Browser-based agentic capabilities moved into the ChatGPT desktop app, a ChatGPT Chrome extension and sidebar, and Codex. Atlas bookmarks, history and open tabs did not transfer automatically, and cookies or signed-in sessions cannot be imported into another browser.
A browser agent is an AI that operates a real browser on your behalf: it opens pages, clicks, scrolls, fills forms and reads what comes back, instead of only answering from training data or a search index.
OpenAI gave roughly a 30-day wind-down. The announcement landed on 9 July 2026 and the shutdown followed a month later, according to OpenAI's own deprecation notice.
The honest reading: Atlas was macOS-only for its entire eight-month life, which is a large part of why it never built a base worth maintaining.
If you built any repeatable workflow on it, that workflow is now broken, and the migration path is not a like-for-like swap.
It is worth being precise about what OpenAI actually kept. The agentic capability survived. The browser as a product did not.
That distinction matters because it tells you where the category is heading: agents living inside a desktop app or an extension, rather than a separate browser you are asked to adopt wholesale.
What can a browser agent actually do reliably today?
Browser agents are reliable at reading and summarising across many open pages, extracting structured data from listings, comparing options across sites, and filling in low-stakes forms. They are unreliable at anything behind a login wall they have to negotiate, anything involving payment, and any task longer than roughly 15 to 25 steps.
The useful mental model: a browser agent is a fast, tireless, slightly careless intern who cannot ask you a clarifying question mid-task.
Where they earn their keep
--- Competitive scans: open 12 competitor pricing pages, pull plan names, prices and limits into one table.
--- Source gathering: collect 20 URLs on a topic with a one-line summary and a publication date for each.
--- Form-heavy admin: repeat the same submission across a directory or a supplier portal you are already logged into.
--- Verification passes: check whether 30 links in a published article still return a live page.
Where they fall apart
--- Cookie walls, CAPTCHAs and two-factor prompts, which stop the run cold.
--- Sites that render content only after scrolling or interaction the agent does not think to perform.
--- Long chains where an early misread quietly corrupts everything after it.
How do you write a browser agent task that does not fail halfway?
Write the task as a bounded work order, not a wish. Name the exact starting URLs, the exact fields to collect, the output format, a hard step limit, and an explicit instruction on what to do when a page blocks it. Vague instructions are the single largest cause of half-finished browser agent runs.
Most people type something like "research my competitors' pricing" and then wonder why the output is thin. The agent had to invent the scope, and it invented a small one.
Here is a template that works across the ChatGPT desktop app, the Chrome extension, Comet and Claude for Chrome. Copy it and swap the bracketed parts.
Try this prompt
You are operating my browser. Complete this task in at most 20 steps.
TASK: Visit each of these URLs in order and extract the pricing information.
URLs: [paste 5 to 10 exact URLs, one per line]
FOR EACH URL, COLLECT: plan name, monthly price and currency, what the cheapest paid plan includes, any usage limit stated in numbers, and whether a free tier exists.
OUTPUT: a markdown table with one row per plan. Add a final column called "Source URL".
RULES: Do not guess. If a value is not stated on the page, write "not stated". If a page requires a login, a CAPTCHA or a cookie consent you cannot clear, skip it and add it to a list called "Blocked" at the end. Do not click any button that starts a purchase, a trial, or a sign-up. Report the number of steps you used.
The three lines doing the heavy lifting are the step limit, the "not stated" rule, and the "Blocked" list. Together they convert a silent failure into a visible one.
The step limit matters more than it looks. Without one, an agent that hits an unexpected page will keep exploring, burn through its context, and return a summary of its own wandering rather than your data.
The "do not click any button that starts a purchase, a trial, or a sign-up" line is not paranoia either. Agents genuinely do click "Start free trial" when a price is only visible after doing so.
One more refinement once the basic version works: ask for the output twice, once as the table and once as a three-line summary of what was missing. The second output is where you find out how much the agent could not see.
Which browser agent tool should you use now?
For most practitioners the honest answer is the ChatGPT desktop app or the ChatGPT Chrome extension, because they inherit the Atlas capability and need no new subscription. Perplexity Comet is the stronger choice if your work is mostly multi-source research. Developer-grade options such as Browser Use and Firecrawl exist, but they assume you are comfortable writing and running code.
The practical shortlist
--- ChatGPT desktop app: OpenAI's designated successor for deeper agentic browsing, with multiple tabs, downloads, improved navigation and account login support. Best if you already pay for ChatGPT.
--- ChatGPT Chrome extension or sidebar: lighter, sits beside the browser you already use. Best for assisted browsing rather than unattended runs.
--- Perplexity Comet: answer-first browsing extended into a full browser. Best for research sessions that compare many sources at once.
--- Claude for Chrome: an extension-based agent for people already working inside Claude.
--- Browser Use or Firecrawl: open-source and API-first, genuinely powerful, and firmly outside the no-code lane.
Availability of the ChatGPT desktop app and extension depends on your plan, region, device and workspace settings, so confirm before you rebuild a workflow around either.
What breaks when you migrate, and how do you avoid losing work?
Three things break in a browser agent migration: your saved context, your logged-in sessions, and your prompt habits. Bookmarks must be exported as an HTML file and imported manually. Cookies and active sessions cannot be moved between browsers at all. And prompts tuned to one agent's quirks rarely transfer cleanly to another.
If you are still holding Atlas data, the export window has closed, but the lesson generalises to every AI tool you depend on.
The migration checklist worth keeping
--- Export bookmarks to HTML the moment a tool announces a wind-down, then import into Chrome via More, Bookmarks and lists, Import bookmarks and settings.
--- Copy open tab URLs into a plain document. Open tabs almost never transfer.
--- Treat exported cookie and session files as sensitive credentials. Never share them.
--- Keep your agent task prompts in a file outside the tool. The prompt is the asset; the tool is rented.
--- Re-test each workflow on the new tool before you trust its output. Same task, same URLs, compare results by hand once.
That last point is where most people lose a week. A workflow that silently degrades is worse than one that visibly stops.
How do you know if the agent's output is trustworthy?
Check three things: whether every row has a source URL, whether "not stated" appears anywhere, and how many steps the run reported. An agent output with no source URLs, no gaps and no step count is usually a confident summary rather than a genuine browsing run.
Agents hallucinate most convincingly when they cannot reach a page. They will produce a plausible price for a plan they never saw.
The "not stated" instruction in the prompt above exists precisely to make that failure visible. If a 10-URL run comes back with zero gaps and zero blocked pages, be suspicious rather than pleased.
Spot-check two rows manually. It takes 90 seconds and it is the only honest quality gate available to you.
Related reading: our earlier piece on connecting AI to your own tools with MCP servers covers the other half of this problem, giving an agent structured access instead of making it click through a webpage.
What should you do in the next 20 minutes?
Pick one recurring research task you do by hand, gather its exact URLs, and run the template prompt above once. Then spot-check two rows. You will learn more about what your browser agent can and cannot do from that single run than from a week of reading tool comparisons.
Good first candidates: a monthly competitor price check, a link audit on your top 10 pages, or a supplier list refresh.
Save the prompt in a document, not in the chat. When the next tool is deprecated, and one will be, you will migrate in an afternoon instead of a fortnight.
The pattern underneath all of this is simple. Tools get retired on someone else's schedule. The workflow you wrote down, and the judgement you apply to its output, are the parts that survive. We understand AI. We understand you better. With UD by your side, AI doesn't feel cold.
Reviewed by the UD AI team.
Ready to Build a Workflow That Survives the Next Deprecation?
Rebuilding one prompt is easy. Rebuilding a workflow your team depends on is not. UD has spent 28 years helping Hong Kong businesses put technology to work, and we'll walk you through every step, from tool selection and prompt design to a documented workflow that outlives any single vendor.
