Most people arrive here one of two ways: you saw “agent mode” in the ChatGPT tools menu and wondered what it did, or you used it, went looking for it, and found it gone. Both need the same answer — and it is not the answer most articles on this topic are still giving.

Quick answer

What agent mode actually did

The mechanism matters, because it explains both what agent mode was uniquely good at and why the replacements do not cover the same ground.

Selecting agent mode spun up a virtual computer in OpenAI’s cloud with four tools on it: a visual browser that rendered pages and interacted with them through screenshots, a lighter text browser for fast reading, a terminal for running code, and direct API access. It could also read connected apps such as Gmail, Google Calendar and GitHub for context.

The visual browser is the important part. Because agent mode literally looked at rendered pages and clicked them, it could operate interfaces built for humans: dropdowns, product configurators, country selectors, multi-step checkouts. Text-scraping cannot do this. And when it hit a login wall it paused and handed you the browser — you typed the password yourself with screenshots switched off, then gave control back and it resumed. That handoff is what made logged-in workflows possible at all.

A typical run took five to thirty minutes. You could interrupt, redirect or stop it and take partial results. It asked permission before consequential actions and refused high-risk ones outright, such as bank transfers.

What worked, and what never did

Even at its best, agent mode had a clear competence boundary — and the same boundary applies to every agentic browsing product on the market.

This worked

• Structured comparison research. Visiting eight vendor sites, pulling pricing and feature data, returning a populated spreadsheet.
• Spreadsheet work. On SpreadsheetBench it scored 45.5% when allowed to edit files directly, against 20.0% for Copilot in Excel. Humans scored 71.3% on the same set — so, better than the alternatives, well short of a person.
• Data wrangling. Taking a messy export, cleaning it in the terminal, handing back something usable.
• Reading behind a login you opened for it — your own order history, dashboard figures or account settings. It also set a state-of-the-art 68.9% on BrowseComp, OpenAI’s obscure-information benchmark.

This didn’t

• Anything urgent. Fifty minutes to find a lamp is not a workflow. It was a project runner, not an assistant.
• CAPTCHAs and bot detection. A large share of failures were “prove you’re not a robot” walls it could not pass.
• Calendar and scheduling judgement. Consistently the weakest reported area — time zones and conflicting constraints need reasoning the browsing loop did not reliably apply.
• Reporting accurately on its own work. The most-cited failure: claiming it had added items to a cart when it had only done so inside its throwaway virtual browser, which vanished with the session.

The pattern is consistent: good at bounded, tolerant tasks where a 70%-correct draft saves time, bad wherever a wrong answer costs more than doing it yourself.

What happened in August 2026

OpenAI withdrew agent mode. The help centre article now opens with a notice that it is no longer available, pointing users to ChatGPT Work and the cloud browser instead. Two things make this messier than a routine deprecation.

The documentation contradicts itself. The banner was bolted onto the top of the existing article rather than replacing it. Directly underneath, the same page still explains how to start agent mode (“select it from the tools menu or type /agent in the composer”), still lists availability across Pro, Plus, Business, Enterprise and Edu, and still publishes the old monthly limits. If you followed those instructions and found nothing, the page was wrong, not you. Rechecking on 23 August 2026, the contradiction is still there — and it is not a fresh oversight. The help centre stamps the article as edited seven days earlier, so OpenAI has been back into the page since the notice went up and left both versions standing: the removal banner at the top, and below it the /agent instructions, an availability list covering Pro, Plus, Business, Enterprise and Edu in all supported countries and territories, the 40-, 400- and 40-messages-a-month limits, and the takeover-mode guidance for logins. OpenAI has not published any plan or region where agent mode survives, and that stale availability list is not evidence that it does — read it as documentation nobody deleted. We cannot inspect other people’s accounts, but the reports on OpenAI’s own forums describe the toggle vanishing from Plus and Pro alike, across web, mobile and desktop.

There was no migration path. The change did not appear in the release notes, which did announce the retirement of Atlas, group chats and the DALL·E GPT in the same period. As of 23 August 2026 the release notes still carry no entry for it. Users raised it on OpenAI’s developer forum on 8 August 2026, arguing that Work, Codex, deep research and desktop computer use each cover part of the old surface and none reproduces the virtual-computer-plus-visual-browser workflow. No notice, no deprecation window, no guide.

That is a real cost to real users. It is also the normal shape of a two-year-old category: Operator lasted about six months before being folded into agent mode, and Google shut Project Mariner down in May 2026 for the same reason. Standalone browsing agents keep getting absorbed into bigger assistants.

What replaced it

ChatGPT Work

Launched 9 July 2026, Work is OpenAI’s agent for longer, multi-step tasks. It researches and analyses, works across connected apps and files, and produces finished documents, spreadsheets, presentations, reports and Sites. You can follow its progress, change direction and approve important actions, and run it through Scheduled Tasks — once, on a schedule, on a trigger, or monitoring for changes.

Work is included on every ChatGPT plan from Free upwards — OpenAI’s Codex pricing page lists it under Free, Go, Plus, Pro, Business, Edu and Enterprise — so what the plan buys is the size of the allowance rather than access itself. It runs on web, mobile and the desktop app. On desktop it can also use local files and other desktop apps with your permission, which agent mode never could — a genuine gain. What it does not do is drive a browser through a site you are signed into and hand control back mid-task.

Cloud browser

Cloud browser handles web tasks Work cannot complete through a connected app. It navigates public pages, enters information in supported fields, and combines site actions with data from connected apps — checking restaurant availability, finding stock nearby, matching flights to your dates, tracking a package. Unlike Work, it is restricted to paid plans: OpenAI excludes Free and Go, and says availability may vary by region during rollout.

The limitation is stated plainly in OpenAI’s documentation: at launch, cloud browser works only on public pages. It does not accept credentials, use autofill or password managers, sign in to websites, or complete payments. If a site requires one of those steps, the task stops, and saved cookies do not enable authenticated browsing.

So the honest summary: the most useful thing agent mode did — getting past a sign-in with your help and continuing — currently has no direct equivalent. Test your workflow rather than assuming it carried over.

What it costs

Agent mode was paid-only and metered by task, not by time: 40 messages a month on Plus, 400 on Pro, 40 on Business and Enterprise, with flexible-pricing workspaces charged 30 credits per message. Only the initial request counted — clarifications and authentication steps did not, but each scheduled-task invocation did.

ChatGPT Work uses a different structure, and the honest headline is that there is no monthly message allowance any more. OpenAI’s Codex pricing page states flatly that Work inside ChatGPT uses the same pricing, credits and usage limits as Codex, while warning that its published examples are coding tasks, so Work consumption will vary. Usage is metered in rolling five-hour windows shared between local messages and cloud chats, with additional weekly limits possible on top.

The only per-plan figures OpenAI publishes are estimated local messages per five-hour window. On GPT-5.6 Sol that is roughly 10–100 on Plus, 50–500 on Pro 5x, 200–2,000 on Pro 20x and 10–100 on Business. Those are not Work numbers: OpenAI notes that cloud chats run on Sol and may use more of your allowance than local messages, so treat them as an upper bound rather than a quota. There is no published credit allowance for Work specifically. What is published is the rate card — GPT-5.6 Sol costs 100 credits per million input tokens, 10 per million cached input tokens and 500 per million output tokens, with typical GPT-5.6 usage put at 5 to 30 credits a message. Past the included limits, Plus and Pro users buy credits individually; Business, Edu and Enterprise workspaces on flexible pricing buy workspace credits.

List prices as we checked them on 23 August 2026: Free $0, Go $8 a month, Plus $20 a month, Pro from $100 a month, and Business $20 per seat per month billed annually or $25 billed monthly, with a two-seat minimum. Pro is really two tiers, and the clearest published source is the Codex pricing page, which names them as Pro 5x at $100 a month and Pro 20x at $200 a month; the main ChatGPT pricing page advertises Pro only as a “from” figure.

The practical point: agentic runs are expensive in a way chat is not. A single task consumed a meaningful slice of a monthly allowance, and under the five-hour-window model it can still empty one afternoon’s capacity. Treat these features as a budget you spend deliberately, not a tool you leave running.

How it compares to Claude and Gemini

All three major labs have reached the same conclusion — browsing agents belong inside the assistant, not beside it — and all three have retired a standalone product to get there.

Anthropic’s Claude in Chrome takes the opposite architectural bet: instead of a virtual machine in the cloud, a Chrome extension that operates your browser, in your session, with your logins already active. It reads the page, clicks, types and fills forms while you watch. That side-steps the login problem entirely — but the work happens in front of you, and only in Chrome. After a Max-only pilot from August 2025 it is now in beta across Anthropic’s paid plans.

Google’s Gemini Spark is where this line of work ended up. Project Mariner was discontinued on 4 May 2026, roughly seventeen months after it was first shown, and Google announced Spark at I/O a fortnight later on 19 May 2026; if you remember this product as Gemini Agent, that is the same lineage under an older name. Spark requires a personal Google Account, an over-18 user, a Google AI Pro or Ultra subscription and Keep Activity switched on, and runs in the Gemini mobile app, the Gemini app on Mac and the web app. Google says it is available wherever Gemini Apps are supported except the European Economic Area, Nigeria, Switzerland and the United Kingdom — so most of Europe is excluded outright — and Pro-tier access reached more than 160 additional countries on 30 July 2026. Chrome’s auto browse, the part that actually drives a browser, is narrower still: United States only, desktop Chrome only, device language set to English, personal accounts only, and capped at 20 multi-step tasks a day on AI Pro or 200 on AI Ultra. It is unavailable on iPhone and iPad, in Incognito, and on school accounts. It does, though, do the thing OpenAI’s cloud browser will not — with your permission it can draw on Google Password Manager to sign you in to a site, and it hands the tab back to you for payments and account creation.

If your work happens inside accounts you are signed into, the extension model is currently more reliable. If it needs long unattended runs producing a deliverable, the cloud model fits better. Neither is mature.

The failure modes nobody puts in the demo

These applied to agent mode and apply, in varying degrees, to everything that replaced it.

• It is slow. Five to thirty minutes was typical, longer was common, and any “how long would this take me” comparison has to count supervision time too.
• It gets stuck in loops. Cookie banners, dynamic dropdowns and CAPTCHAs produce retry cycles that burn the run.
• It misreads pages, then reports success. Screenshot-based interaction means ambiguous layouts get read wrongly and the agent proceeds confidently. The gap between “the agent said it did X” and “X happened” is the single most important thing to verify.
• Websites block it. Sites increasingly refuse automated browser traffic, so a page that works in your own browser may be blocked outright. That is the operator’s decision, not a bug you can fix.
• Prompt injection is a live risk. A malicious instruction hidden in a page can try to redirect an agent that has access to your email and files. OpenAI ships confirmations, refusals, monitoring and a supervised watch mode, and is explicit that these reduce but do not eliminate the risk. Google says the same about Spark and auto browse. Enable only the connectors a task needs, and avoid prompts like “check my email and handle everything.”

If your workflow broke

The most common way people discovered the removal was a scheduled task that quietly stopped producing anything. There is no automated migration: open chatgpt.com/schedules, find the recurring tasks that invoked agent mode, check their last few runs, rebuild the ones that matter, and pause the rest.

Decide per task. Research ending in a document goes to Work. Public-web gathering goes to Work or cloud browser. Anything needing a sign-in has to be tested against cloud browser first, and may need a different tool entirely. Code-shaped automation goes to Codex. And if what you wanted was a connected, tool-using assistant rather than a browsing session, the more durable pattern is wiring real tools to the model — see our guide to MCP servers, which solve the same problem through APIs rather than screenshots.

The honest verdict

Agent mode was an impressive demonstration and a mediocre daily tool — genuinely useful for a narrow band of bounded research and spreadsheet work, and for reading behind logins you opened for it. Its removal cost users a real capability without a replacement, and OpenAI communicated it badly. But the lesson is not “the agent era is over”; it is that this category is still churning hard, one year in, and anything you build on it should be rebuildable. For the longer arc, see the AI revolution: from bots to intelligent agents.