
ChatGPT agent was a ChatGPT mode that carried out multi-step tasks on its own virtual computer, browsing websites, running code, and building files, before OpenAI retired it in favor of ChatGPT Work.
OpenAI launched it in July 2025. You described the outcome, and it handled the steps. TechCrunch reported at launch that it could complete a wide variety of computer-based tasks. If you cannot find agent mode now, nothing is wrong with your account.
What ChatGPT agent was
Before agent mode, ChatGPT answered, and you did the work. Agent mode changed who did the doing: you handed over the job, and ChatGPT worked through the steps on its own cloud computer, narrating as it went.
It combined Operator, which scrolled, clicked, and typed on websites; deep research, which read many sources into one report; and ordinary chat, so you could interrupt or take over. One launch guide describes it as ChatGPT getting its own computer on OpenAI’s servers.
OpenAI called it early and said it could make mistakes. The ChatGPT models guide explains which models power ChatGPT now.

How it worked on its own virtual computer
On that computer it had a visual browser that saw pages as a person does, a text browser that read pages as plain text, and a terminal for code and files. It picked whichever fit each step, from skimming pages as text to reworking spreadsheets.
Connectors to apps such as Gmail and Google Calendar let a task start from your inbox or calendar, and files went both ways: a document in, a slide deck or spreadsheet back, with source links attached.
Logins used takeover mode: the agent paused at a sign-in page while you typed your own credentials, capturing no screenshots. It also asked before consequential actions, such as sending an email or making a purchase. OpenAI’s help page lists the safeguards, including confirmations for high-impact actions and watch mode.

What replaced it: ChatGPT Work and cloud browser
Agent mode began disappearing in July, a year after launch, and OpenAI’s help page now says ChatGPT agent is no longer available. BytePicks checked the help pages: OpenAI points users to ChatGPT Work for longer tasks and cloud browser for website tasks. The old instructions still sit further down that page, so many guides still call it live.
Work is the agent side of ChatGPT now, built for longer, multi-step work and finished deliverables. You choose Work instead of Chat when you start a task. Cloud browser is Work’s browser, on a separate computer in the cloud, and ChatGPT decides when a task needs it. By default it asks before visiting a new website, paying, or booking.
Both are paid features: cloud browser is not on Free or Go, and sign-in support is documented for Plus and Pro. Plus costs $20 a month; Pro has three tiers at $100, $200, and $500 a month, in US dollars. Confirm prices before you pay. The old limits, 40 messages a month on Plus and 400 on Pro, went away with agent mode; Work usage is counted like Codex usage. The free AI tools for students roundup lists no-cost options; the best AI tools compared puts ChatGPT beside its rivals.

A worked example: research, compare, ask before booking
Say you need a hotel for a three-night work trip. You set the outcome and limits: city, dates, budget, and a clear line that nothing gets booked without your say-so.
The agent searched the web, opened the booking pages, and read live prices and terms instead of answering from memory, then built a spreadsheet comparing price, distance, and cancellation terms. Then it stopped: booking meant a payment, so it asked for your choice and, if the site needed your account, for you to sign in.
That pause is the point: the agent did the slow part, and you kept the decision that cost money. If a site threw up a CAPTCHA or bot check, the agent could stall, and you finished the booking yourself. The ChatGPT alternatives guide compares other tools for the same kind of task.

What to delegate and what to keep for yourself
Agents earn their keep on structured, checkable work: research that ends in a table, first drafts from sources you can open, price gathering, and summaries of long reports. You can verify each result in minutes, and a mistake costs little.
Keep the rest: purchases, passwords, and anything you cannot undo. Vague instructions are a trap too. “Check my email and handle everything” hands the agent your inbox with no boundaries, which OpenAI’s guidance warns against. Give it a narrow job, named sites, and a clear stopping point.
Two limits matter: tasks often took many minutes, and sites with logins, CAPTCHAs, or bot checks could end one early. A Litmus Tools review found most tasks ran 5 to 30 minutes and CAPTCHA-walled pages routinely stopped the agent. Stay nearby while it works. The Claude Code guide covers an agent built for software work; the DeepSeek AI guide covers another model for research.

The prompt injection risk
The sharpest risk with any browsing agent is prompt injection: hidden instructions in a web page, email, or document that steer the agent into something you never asked. A model reads all text as text; it cannot reliably tell your instructions from a stranger’s.
Take OpenAI’s own example. You ask it to find a restaurant by checking your calendar and emails, and it meets a malicious comment directing it to fetch a password reset code from Gmail and send it to an outside website. OpenAI built in confirmations, refusal patterns, and injection monitoring, but says the risk remains. The UK National Cyber Security Centre has warned it may never be fully mitigated.
Enable only the apps a task needs, keep ask-before-visiting on, and read every confirmation request before approving it. Never type a password or security code into the chat; use the secure sign-in flow. Clear the remote browser’s data after sensitive sessions, and stop any task that looks off.

Frequently asked questions about ChatGPT agent
Is ChatGPT agent still available?
No. OpenAI’s help page says it is no longer available. Use ChatGPT Work for longer tasks and cloud browser for website tasks.
What replaced ChatGPT agent mode?
ChatGPT Work, for longer tasks and finished deliverables, with a cloud browser for the website steps.
How much does ChatGPT Work cost?
Work has no separate fee, but on the web and mobile it needs Plus ($20 a month) or Pro ($100, $200, or $500 a month). Cloud browser is not on Free or Go.
Was ChatGPT agent available in Europe?
It varied by region and plan, and the same is true of Work now. If you are in the EEA or Switzerland, check whether Work appears in your account.
What is prompt injection?
Hidden instructions in content an agent reads that steer it into actions you never asked for. Specific instructions, few connected apps, and careful confirmations reduce the risk.
Can an AI agent book or buy things on its own?
It can prepare everything up to the payment, but it must ask before a booking or purchase. Never approve a payment you have not read.