ChatGPT agent is retired. Gemini tasks skip the EEA and the UK.

ChatGPT agent is gone. The personal assistant that replaced it still stops at the payment. On November 14, 2026, OpenAI’s help page titled ChatGPT agent says the agent is no longer available and points you to ChatGPT Work. Work’s cloud browser, on paid plans except Free and Go, can book a DMV slot only as far as “prepare it for your approval,” and it asks before a payment.

Gemini tasks are the other hard limit, and it is geographic. Google’s tasks article says they run wherever Gemini Apps are supported, except the European Economic Area, Nigeria, Switzerland, and the United Kingdom. They also refuse a work or school Google account. Claude Cowork will do the file and the inbox, then the safety page leaves the purchase on you.

A personal assistant AI, on this page, means a product that takes a step on one person’s behalf: a browser, a file, a message, a form. It is not a framework you import. The library decision is a different note, which of 11 frameworks to install, counted the same morning. An agent is a model plus the harness that runs tools and stops. Autonomy is how far that loop goes before a person confirms the next step. The confirm is the whole subject here.

What you can still hand off this week

Three products, three different computers. ChatGPT Work uses a browser OpenAI runs. Claude Cowork runs the task on Anthropic’s servers and, if you turn it on, can also click your screen. Gemini tasks can use a remote browser or the Chrome already open on your desk. The stop rule is the column that should decide the handoff.

The prices of the plans are not in this table. The help pages used here name the plan families and do not print a monthly dollar amount for Plus, Pro, or Google AI Pro. This page does not borrow a price from a different article.

ChatGPT Work keeps going after you close the laptop

The agent article and the Work article disagree with each other, and the agent article disagrees with itself.

The ChatGPT agent article, read November 14, opens with the current instruction: ChatGPT agent is no longer available. Use ChatGPT Work for longer, multi-step tasks. For the browser, use the cloud browser article. The help center marked that agent article “Updated: 17 days ago,” which on this read is late October. The rest of the article still documents the retired mode. It lists Plus at 40 messages a month, Pro at 400, and Business and Enterprise at 40, and it still explains Operator as something that was folded into agent mode. Do not staff a Work rollout from those caps. They describe a product the first paragraph says you cannot open.

The Work article, marked “Updated: 8 days ago” on the same read, is the live one. Work researches, analyzes, and creates a document, a spreadsheet, a presentation, a report, or a Site. It runs once, on a schedule, or on a trigger. On web and mobile it runs in the cloud. In the desktop app it can use local files and desktop apps if you allow that, and the article says messages and task context may still be stored in the cloud when the work itself is local. Web and mobile Work cannot see the files on your computer.

Event-triggered tasks, for eligible Plus, Pro, Business, Enterprise, Edu, and ChatGPT for Healthcare users, fire on a new Gmail message, a Slack channel message, or GitHub pull-request activity in a repository you authorized. They are not available on Free or Go, or in FedRAMP workspaces. For Slack you add @ChatGPT to each channel the task watches. Actions that need approval still need it. On web and mobile, spoken approval is not supported. You approve on the screen.

The cloud browser article, also marked 17 days before this read, is the part that acts on the web. The browser sits on a separate computer. It can read pages, click, and fill forms, on public sites and on sites where you sign in. You can start on web or mobile and leave. It pauses for input, sign-in, or confirmation. It is on paid plans in supported regions, excluding Free and Go. It does not use the cookies or the saved passwords in the browser on your desk. You sign in again, in a secure form. A review model checks that form for phishing before you type. The username and password go to the remote browser. The article says they are not visible to the model, and that ChatGPT does not store them.

The examples are specific, and the last step is often yours. Check a restaurant, compare flights, track a package from a connected inbox, find a DMV appointment and prepare the booking for your approval, compare utility plans after you sign in, save apartment listings, reconcile invoices. The article says some tasks need your review or a final step you complete. Allowing a website does not allow the consequential action. ChatGPT is designed to ask before a booking or a payment, and before anything that is hard to reverse or that creates a financial, legal, or account commitment.

That ask is the approval gate. The tool use is the browser or the connector. The gate is the chat confirmation. Turning a site to “always allow” does not remove the gate. The article says so.

Wallet is not this gate. ChatGPT Wallet, opened November 13, pays from a network token after Face ID or Touch ID, and it runs in a dot, not in ordinary Chat and not in Work. Poppy, Sierra’s Personal Agent Protocol 1.0, is how a company tells a visiting agent what the person allowed. Both are the November 14 Wallet explainer. This page does not restate the card rules.

Claude Cowork runs in the cloud. The purchase stays yours

Anthropic’s safety article, read November 14, splits the tools in a way the other two vendors only imply. Read tools look: an inbox, a screenshot. Write tools act: a calendar invite, a deleted file, a command, a click. Write tools are where a mistake becomes a fact. The article says human oversight is the recommendation when the stakes are high, because Claude can be wrong.

The task itself runs in the cloud, in an isolated environment that exists for that session, cannot reach your home or company network, and is removed when the session ends. Isolation limits where the code runs. It does not limit what Claude can read if you connected the inbox. A prompt injection, in the article’s example, is a malicious instruction sitting in a message you asked Claude to summarize. The attack needs both a read outside your trust boundary and a write that can hurt you. Close one of those and the attack fails. That is the practical rule, not a slogan.

Cowork on the desktop is on every paid plan the article names: Pro, Max, Ultra, Team, and Enterprise. On web and mobile it is on Pro, Max, Ultra, and Team, with Team in beta, and on Enterprise when an owner turns it on.

Computer use is narrower, and it is the one people confuse with “it can use my computer while I sleep.” It is a beta for Pro, Max, and Ultra. It is in the desktop app on macOS and Windows. Team and Enterprise do not have it. The order is connectors, then a browser, then the screen. Gmail, Drive, Microsoft 365, or Slack, if connected, beat a click, because a click is slower and more likely to miss. Claude asks before each application. Some apps are off by default. The computer has to be awake, and the desktop app has to be open. On macOS 15 or later it can work in background windows and leave your pointer alone. That is still your machine, powered on. It is not the cloud task.

The safety article is blunt about the schedule. Scheduled tasks run when you are away, including when the computer is off. Do not schedule a task that sends a message, makes a purchase, touches sensitive files, or does anything hard to undo. Review the runs. Pause the ones you are not using. And the responsibility line is not a suggestion: you remain responsible for content Claude publishes, messages it sends, purchases and other financial transactions, and actions taken through computer use. The click loop, as a category rather than as Claude’s current beta, is Can an AI agent use your computer?. This page does not edit that guide. The rules above are the help center on November 14.

Gemini tasks stop at a work account, and at four places

Google’s tasks article, read November 14, is a checklist before it is a feature list. You have to be 18 or over. You sign in with a personal Google account. A work or school account cannot use tasks. You need Google AI Pro or Ultra. Keep Activity has to be on. Tasks then run wherever Gemini Apps are supported, except the European Economic Area, Nigeria, Switzerland, and the United Kingdom. The surfaces are the Gemini mobile app, the Gemini app on Mac, and gemini.google.com.

Inside those limits the examples are ordinary and useful. Declutter an inbox by summarizing or archiving newsletters. Build a to-do list after a vacation. Unsubscribe from lists. Watch flight and hotel prices and alert you when they drop inside a budget. Pull cited research into a file. The article also says you can ask for tasks that shop, book travel, or reserve a table. “Can ask” is not “will pay.” The safety section says Gemini is designed to ask every time before it sends a communication, modifies or deletes your data, makes a purchase, completes a financial transaction, or fills out and submits a form. It also asks before Sign in with Google. Passwords and payment details are a “take control” pause: you type them on the page, not into the task thread. The article says not to put sign-in info or payment details in the thread at all.

The browser split matters if you leave the desk. Local Chrome can see the sites you are signed into, and the device and Chrome have to stay awake. If you close the device, Gemini might use a remote browser to finish. Auto browse, the Chrome feature, is a further gate. The auto browse article requires you to be 18 or over and in the US, on a personal account with Google AI Ultra, Pro, or Plus, with the device language in English. A school account in Chrome cannot use it. A work or school account needs an administrator. The first browse asks Chrome for permission. After that, every browsing task asks you to confirm before it starts. Gemini may share your name, contact information, files, and preferences with the sites it chooses. The article says the safeguards do not guarantee protection, and that your supervision is the control that matters.

There is one action that does not ask. The tasks article says Gemini can perform bulk actions on private tasks in Google Tasks without your confirmation, and it tells you to review the request before you submit, or to manage those tasks in Google Tasks yourself. A purchase waits. A stack of private to-dos does not. If you were going to trust one unattended write, do not make it that one without reading the list you are about to hand over.

A scheduled task has a second gap. If it runs while you are offline, you may not be able to stop an action you did not intend. Some tasks still require confirmation. The article does not promise that every schedule will wait.

Gemini 4 model prices are a different page, Hydrogen, Carbon, and Argon, read November 13. This note is the product that acts, not the token rate. The morning’s fetch brief, GoogleAgent after Gemini 4 Hydrogen, counts this site’s own answer fetches. It is not a review of Gemini tasks.

What to trust, and what to keep

Trust the draft. Trust the research packet. Trust the price watch that only alerts you. Trust a form that stops on the confirmation screen.

Keep the payment, the send, the delete, and the password. ChatGPT’s cloud browser is designed to ask before the booking and the payment, and the password never enters the model. Claude’s safety page puts the purchase on you even when the product asks, and it tells you not to schedule one. Gemini is designed to ask before the purchase and the submitted form, and it tells you to type the card on the page yourself. None of those sentences is a guarantee. Each vendor says the safeguards fail sometimes. The approval gate is a pause in the product. It is not a refund.

Keep a work account out of Gemini tasks. The article does not offer a setting that turns them on. Keep a Team or Enterprise Claude seat off computer use, because the beta is not on those plans. Keep a Free or Go ChatGPT account off the cloud browser, because the article excludes them. Keep the EEA, the UK, Switzerland, and Nigeria off any plan that assumes Gemini tasks exist there.

For a builder wiring an agent of your own, the consumer stop is the spec to copy. Put the spend, the send, and the delete behind a confirmation the model cannot click through. Do not treat “always allow this site” as permission to pay. Do not put a card number in the tool result. The framework that checkpoints the pause, if you are building rather than buying, is the November 14 install note. The shopping agents that can finish a checkout on a merchant’s own rails are the six-month shopping report, not these three personal products.

Where the other guides start

This page stops at the help center. The payment products are already written.

• Wallet, the dot, and Poppy 1.0: How does ChatGPT Wallet work?
• Which assistant can finish a store’s checkout: Can an AI shopping agent check out for you?
• The desktop click loop: Can an AI agent use your computer?
• How a browser agent reads a page and stops: How do AI browser agents complete tasks?
• The year ahead, including what is already dated: AI agents, the next 12 months