Can the assistant use a web browser?

Yes. The assistant can open a real web browser and work a website for you: going to a page, clicking, typing, filling in fields, scrolling, reading what is on screen, and waiting when a page is slow. You watch it happen live in a panel beside the chat, and you can take the controls at any moment.

The feature is called Computer use, and it lives inside the assistant chat. There is no separate page to go to and no setting to switch on first. You ask for something, and if the job needs a browser, a browser opens.

When it uses a browser, and when it just answers

The assistant applies a simple test. Could a person get this done by reading a page? Then it reads and answers in the normal way. Would a person have to work the site — click, type, filter, sort, sign in, page through results, fill something in? Then it opens a browser.

So a fact, a definition, today's news, or a summary of a page it can simply fetch comes back as an ordinary answer. That is deliberate: the plain answer is quicker, so it comes first whenever it genuinely does the job.

These are the kinds of request that send it to a browser:

• Book, order, reserve, apply, submit or register for something.
• Check your account, your balance, an order, a booking, a tracking number or an invoice.
• Find flights, hotels, tickets, appointments, listings or prices to compare.
• Fill in a form.
• Look through a portal, dashboard or inbox that has no other way in.
• Download something that sits behind a login.
• "Can you sort out X for me", where X lives on a website.

It will also reach for a browser when an ordinary search or page fetch came back thin, out of date, blocked, or clearly needed the page to load properly first. If an answer looks useless, telling it to go and actually look at the site is enough. You never have to say the words "browser" or "computer" — describe the job and let it decide.

What you see while it works

The moment a browser task starts, a panel headed Computer slides in from the right-hand edge of the chat on its own. You do not have to open it. Inside it is the browser itself, moving, live.

A card also appears in the conversation for that run, with a Watch button. While the task is going, both the status label in the panel header and the card headline narrate the step it is on rather than just saying it is busy — "Opening" and the address, "Reading the page", "Clicking", "Typing", "Filling in a field", "Waiting". Underneath, the card keeps a running count of actions and shows the address the browser is on.

While a run is going, the chat itself waits: the send button is replaced by a stop button, and the message box reads "Working in the browser…". You do not need to sit and watch, though. The task carries on by itself, and the assistant comes back into the chat with the outcome as a new message, without you asking for it.

It is your browser, and it stays open

The browser belongs to you rather than to the task. When a task finishes it does not close — it is deliberately left open on the page it reached, so you can carry on from there or ask for the next thing. The card reads "Done — browser still open".

Whatever you ask next continues in that same browser: same tabs, same history, same sign-ins. That holds across conversations too, so a brand-new chat carries on with the same computer.

After a stretch with nobody using it, the browser puts itself to sleep. Asleep is not closed: your sign-ins, cookies, site preferences and the page it was on are all kept, and it wakes up where you left it, still signed in. Usually there is nothing for you to do about it — ask for the next thing and the assistant wakes it itself.

What it will not do on its own

Some steps it stops at and hands to you. These are enforced underneath the assistant, so asking again in different words will not get past them. None of them counts as a failure, and none of them closes the browser: it stays open on the page it reached, tells you what it stopped at, and waits while you do that bit yourself.

• It will not type a password, a passcode, or a security, one-time or two-step code.
• It will not enter payment card numbers.
• If it cannot tell what a field is for, it refuses to type into it at all rather than guess.
• It never attempts a "confirm you're human" check.
• It stops before an action that would change something — buying, paying, submitting, sending, deleting, cancelling and the like — unless that is exactly what you asked it to do. It will fill a form in and stop short of sending it, and that includes pressing Enter where Enter would submit.
• If you scoped the task to particular sites, going anywhere else is refused, including a page the site itself redirects to. It names the sites it is allowed on and the one it stopped at, and invites you to add it.
• It opens ordinary web pages only, and it does not run code inside them — it works a site with its ordinary controls, the same way a person does.
• If a page contains text written to look like an instruction to the assistant, it is treated as part of the page rather than a request from you. It does not act on it, and it tells you it saw it.

What it costs

A browser task uses credits while it works. The time the browser is open and the assistant's work driving it are billed together as one charge — not one per click, and nothing extra stacked on top.

The meter runs for as long as the browser is open. That includes after a task has finished and the browser is sitting on the page it reached, and it includes the time you spend driving it by hand. It stops when the browser actually closes: you press Stop, or it goes to sleep. Asleep costs nothing and loses nothing. Waking it opens a fresh browser, so the meter starts again.

It appears on your usage breakdown as its own line, "Computer use".

Tip: if you have finished with the browser, pressing Stop closes it there and then, rather than waiting for it to fall asleep on its own.

Limits worth knowing

A browser cannot run forever. There is a limit on how long a single task may take, so a long job can end with the browser having run out of time before the task was finished. A run can also stop on its own when it reaches a spending ceiling for that run; if it does, the browser is left open on the page it reached and the assistant tells you what still needs doing.

You cannot start a second browser task while one is running. Ask, and the assistant tells you it is still working in the browser and to come back once that has finished — and when you do, it carries straight on in the same browser.

Scheduled automations cannot do this

Computer use is assistant-only. Scheduled and unattended automations cannot drive a browser at all, cannot see which sites you have saved sign-ins for, and cannot start a sign-in. If you ask one to do browser work, it says so in what it hands back rather than failing quietly.

The reason is straightforward: an unattended run has nobody there to clear a "confirm you're human" check, sign in when a site asks, or notice that a page has changed shape since the job was set up — which is most of what a real browser session runs into.

The browser always runs as you, the account owner, and where you have a saved sign-in, as you already logged in. That is also why it is never reachable by someone outside your account talking to the assistant, such as an inbound caller or an inbound email conversation.

Note: if the assistant tells you the browser service is not set up on your account yet, that is what it means — nothing has gone wrong with your request.