ChatGPT Work: The Complete Guide 📚

These two are the "wait, it can do that?" features. Both are desktop only.

The built-in browser

The new desktop app has a real browser inside it. Not a scraper. An actual browser with tabs, sign-ins, autofill, password management, extensions, and downloads. Open it from the toolbar or hit Cmd+Shift+B on Mac, Ctrl+Shift+B on Windows.

You and ChatGPT see the same page. It works across multiple tabs in one task, downloads files, and waits while you sign in to something. There is an annotation mode where you point at an element on a page, leave a comment, and ask ChatGPT to fix it.

It uses its own browser state, not your Chrome profile. If a task needs your already signed-in Chrome session, your open tabs, or your Chrome extensions, use the updated Chrome extension instead, which now puts ChatGPT in Chrome's sidebar.

Computer Use

This is the wild one. Computer Use lets ChatGPT use your computer on your behalf. It clicks. It types. It opens apps. It moves files where they need to go. It runs in the background across your apps, tools, and browser while you do something else.

What it is for: the stuff that has no plugin. The internal dashboard nobody built an integration for. The clunky desktop app your company still uses. The tool that only works if a human clicks through it.

Sites is in public beta and it is the sleeper feature of this launch. You can turn your work into an interactive site or web app and share it with a URL. Not a doc. Not a PDF. An actual live page.

What people are building with it:

• Live dashboards
• Project trackers
• Launch calendars
• Prototypes and internal portals
• Interactive reports

And the good part: ChatGPT can keep them updated as the underlying information changes. Build a dashboard once, connect it to the tools the data lives in, and it refreshes itself.

Build me a Site that shows this week's status: open tasks by owner from @Asana, anything I was tagged in on @Slack, and this week's calendar.

Keep it simple and readable. Refresh the data every time the page is opened.

The reframe: if you have ever built a Google Sheet that four people check every Monday, that is a Site now.

Scheduled Tasks let you ask ChatGPT to do something once, on a repeating schedule, when an event happens, or continuously as a monitor. It runs whether you are online or not. You get a push notification or an email when it is done.

Real examples that actually work:

• Review new Slack updates each week and refresh a recurring meeting agenda.
• Check websites and dashboards every morning, summarize what changed, send a report.
• Monitor new customer feedback and turn recurring themes into prioritized product ideas.
• Update a presentation when new feedback arrives by email.
• Turn new Teams and Slack messages into updated docs or slides, then share the important changes with your team.

How to set one up

Just ask.

Every Monday at 8am, pull my calendar for the week, check Slack for anything I was tagged in, and give me a one-page brief. Lead with anything that needs a decision from me.

Manage all of them from the Tasks page (profile icon, then Tasks), where you can edit the name, the instructions, and the schedule, or pause and delete them.

A project keeps related conversations, files, and instructions in one place. On desktop you can tie a project to a local folder. The point is context. Instead of re-explaining your client, your product, and your format in every single task, you set it once at the project level and every task inside it inherits that.

If you work on the same handful of things week to week (and everyone does), set up a project per thing. It takes five minutes and it saves you from repeating yourself forever.

Do this today. Make one project for the thing you work on most. Put your standing instructions in it: who it is for, what format you want, what to never do. Every task you run inside it gets smarter for free.

Enough theory. The honest test for whether something belongs here: is it high-effort, repeatable, and does it end in a file or a decision? If yes, delegate it. If it is a quick question, just use Chat.

If you run a business

1. Month-end close. Find the source data across your tools, move it into Excel or Sheets, reconcile it, build the slides, check the results. OpenAI's own finance team says this went from days to hours.
2. Competitor teardown. Give it a list of competitors and what to compare. Get back a full dataset showing where you lead and where you lag.
3. Lead triage. Trace every lead across the CRM, call recordings, and email, find where follow-ups broke down, and roll it into a weekly exec dashboard.
4. The monthly board deck. Built automatically from the tools the numbers already live in.
5. Customer feedback into a roadmap. Monitor incoming feedback, cluster the themes, rank them, keep the list live.
6. Invoice and receipt chaos. Point it at the folder. Get a clean expense spreadsheet.

If you are a marketer or creator

1. Research to brief to assets. Customer research becomes a campaign brief becomes the actual assets, in one request, with context carried through every step.
2. Adapt one campaign for five markets without rewriting anything from scratch.
3. Content calendar as a live Site that the whole team can check instead of asking you.
4. Long transcript into everything. Turn one podcast or video into a newsletter, five social posts, and a blog outline.
5. Weekly performance report pulled from your analytics, on a schedule, in your inbox before you wake up.
6. Landing page research. Have it open the built-in browser, look at 20 competitor pages, and tell you what the actual patterns are.

If you work with data

1. Clean up the ugly spreadsheet. Malformed rows, wrong headers, junk data. It fixes it and hands you a real file.
2. Real statistical work. Outlier detection, cross-tabs, time-series, with charts.
3. A dashboard that refreshes itself, built as a Site, pulling from your connected tools every time it opens.
4. Reconcile two systems that disagree and tell you why they disagree.

If you are in sales or ops

1. Meeting prep. Pull the account history, the last three emails, the Slack thread, and the open tickets, and hand you a one-pager before the call.
2. A live account command center that updates itself as activity changes.
3. Launch readiness checks. Automate the checklist and free up your PMs.
4. Event ops. Rebuild your whole conference workflow so pre-event number crunching runs on a schedule.

If you are just a person with a computer

1. Organize the disaster folder. Downloads, Desktop, wherever. Sorted, renamed, deduplicated.
2. Batch rename files with a consistent date format so you can actually find things.
3. Read the 60-page contract and tell you what is unusual in it.
4. Voice memo to polished document. Turn a ramble and some scattered notes into something you can send.
5. A morning brief. Your calendar, your important emails, and whatever you asked it to monitor, waiting for you every day at 7am.

Five of my favorite ways to actually put Work to work. Each one runs on the same connectors, scheduled tasks, and local folders you set up above. Connect the tool, paste the prompt, and let it run.

1. Price Tracking

Give Work your wishlist and it becomes a full-time deal-watcher. It checks the price of every item across every store and pings you the second one drops. Set it as a scheduled task so it runs on its own.

Here's my wishlist: [list your items, with links if you have them]. Check the current price of each one across the major retailers and set this to run every day. The moment any item drops below what I'd pay or below [target price], message me with the item, the new price, and the link.

2. Meeting Prep

Connect your calendar and wake up ready for every meeting. Overnight, Work pulls who you're meeting, researches them and their company, and builds a brief for each one.

Every night at 10pm, look at tomorrow's calendar. For each meeting, tell me who I'm meeting with, research them and their company, and build a one-page brief: their background, recent news, what they likely want, and three smart questions I can ask. Have it ready before I wake up.

3. Refund Hunting

Upload your recent orders and Work goes looking for money you're owed. It reads each store's return and price-adjustment policies, finds what qualifies, and drafts the exact requests to send.

Here are my recent orders [upload receipts or order list]. Check each store's return window and price-adjustment / price-protection policy. Find anything I'm still owed money on (price drops, late deliveries, missing items) and draft the exact refund or price-adjustment request for each one, ready for me to send.

4. Weekly Reporting

Set it once and every Friday you get a full recap. Work pulls from your connected apps, builds the report, and drops it into a folder on your computer automatically.

Every Friday at 4pm, pull this week's activity from my connected apps (Slack, Gmail, Calendar, analytics, whatever's relevant), build a clean one-page report of what happened and what's still outstanding, and save it as a file in [folder] on my computer, named with the date.

5. Travel Logistics

Hand it your flight confirmation and it plans your entire travel day, right down to beating jet lag.

Here's my flight confirmation [paste or upload it]. Build my full travel day: when to leave for the airport, my terminal and gate, which lounges I can get into, security and boarding timing, and a simple sleep schedule to beat jet lag for the destination time zone.

Here is the part the launch coverage buried, and you need to know it. ChatGPT Work does not work like a normal subscription feature.

OpenAI is explicit about it. Work is designed for longer, more involved work than a typical chat request, so usage works differently. Usage varies with the amount of work a task requires, and more complex tasks use more of your plan's included usage. It follows the same usage structure as Codex.

Translation: an agent that runs for two hours across twelve tools burns a lot more of your allowance than asking it to write a haiku. Your subscription buys an allowance, and ambitious agent work meters against it.

For teams, Enterprise and Edu admins get spend controls in the Admin Console. They can set workspace defaults, set group limits, create individual overrides for people who need more capacity, and review requests for extra credits.

ChatGPT Work reads your stuff and takes real actions in the real world. That combination is powerful and it is also where things go wrong.

What OpenAI built in

• Plan mode, check-ins, and approvals. You decide what it can access, when it checks in, and what needs your sign-off.
• Auto-review. OpenAI's most advanced models review important actions involving connected tools before they happen. In adversarial red teaming, OpenAI reports auto-review blocked 100% of attempts to extract protected data.
• Enterprise foundation. Work is built on ChatGPT Enterprise's security, privacy, and compliance base, with a Compliance API for oversight at scale.
• Website permissions. The built-in browser asks before using a new site unless you have already allowed it.

What you should do

• Use a dedicated working folder instead of granting access to everything.
• Keep approvals on for money, sent messages, and irreversible changes.
• Do not point Computer Use at sensitive apps. Banking, health, anything personal.
• Treat web content as untrusted. A malicious instruction hidden in a webpage or an email is a real attack (prompt injection) and it is the main risk with any agent that reads the internet.
• Watch for scope creep. If it starts touching files or sites you did not mention, stop it.

You are responsible for what it does on your behalf. That is not a legal disclaimer. That is just true.

You have probably heard of the other one. Anthropic shipped Claude Cowork earlier this year and it is the same basic idea: an agent that does knowledge work and hands you finished files. I use both. Here is the honest take.

Where ChatGPT Work wins

• Distribution. It is inside the app hundreds of millions of people already open. Zero new habits. That matters more than any feature.
• Sites. Turning work into a live, shareable, self-updating web page with a URL is genuinely ahead. Claude's live artifacts are similar but desktop-only.
• The built-in browser. A real browser inside the app, with tabs, sign-ins, extensions, and annotation mode, beats bolting onto Chrome.
• Free on desktop. Chat, Work, and Codex on every desktop plan including Free. Claude Cowork is paid plans only.

Where Claude Cowork wins

• Sessions run in the cloud and keep going. Close your laptop and Claude keeps working. Your sessions and files follow your account across desktop, web, and mobile.
• Scheduled tasks with no computer at all. They run remotely with no device online.
• Sub-agents. Claude spins up multiple agents working in parallel on different pieces of one job.
• Deletion protection. Claude requires explicit permission before permanently deleting any file, in every mode.
• Writing quality. Subjective, but Claude still writes better prose. For anything with a voice, I reach for Claude.
• Project memory. Cowork projects remember how you like things done within that workspace.

So which one should you use? Honestly, whichever one is already in your life. They are about 85% the same product. Pick the one whose app you already have open, learn it properly, and get on with your life. The gap between "uses an agent well" and "does not use one at all" is a hundred times bigger than the gap between these two.

1. Treating it like chat. Writing a question instead of a brief. Give it an outcome, sources, constraints, and how you will judge it.
2. Not connecting anything. ChatGPT Work with no plugins is just ChatGPT with extra steps.
3. Using web when they need desktop. Local files, the browser, and Computer Use are desktop only.
4. Starting with the hardest thing they have. Start with a task you already know cold so you can actually judge the output.
5. Turning off approvals immediately. The approvals are the product working, not the product being annoying.
6. Scheduling before watching it run once. Run it manually, twice, then schedule it.
7. Ignoring usage. It is metered. Measure your first week before you automate anything.
8. Skipping skills. If you find yourself explaining the same preferences a third time, stop and turn it into a skill.

Do I need to pay? For the desktop app, no. Chat, Work, and Codex are on every desktop plan including Free. On web and mobile it started with Pro, Enterprise, and Edu, then expanded to Plus and Business.

I do not see Work anywhere. The rollout is gradual. Update your app and check again. If you are on the old desktop app, you need the new one.

What is ChatGPT Classic? The old desktop app. It still works and still gets updates, but the new agent features only live in the new app.

Can it see files on my computer? On desktop, yes, for the folders you give it access to. On web and mobile, no.

Is Codex going away? No. It is now a mode inside the new desktop app, and it got better: inline diff editing, PR review in the side panel, faster computer use, multi-repo projects.

What happened to Atlas? OpenAI is sunsetting the standalone browser. Its capabilities moved into the desktop app's built-in browser and an updated Chrome sidebar extension.

What model does it run on? GPT-5.6, which launched the same day.

Can it work while my laptop is closed? Cloud tasks and scheduled tasks keep running. Anything needing your local files, desktop apps, or Computer Use needs your computer awake and the app open.

Is my data safe? Work is built on ChatGPT Enterprise's security and compliance foundation, with an auto-review layer checking important actions before they happen. That said, you are the last line of defense. Do not give it access to things you would not give a new contractor on day one.

For three years, using AI at work meant a chat window sitting next to the work.

As of July 9, 2026, on the most-used AI product on earth, it means delegating the work itself and getting finished files back.

That is the whole shift. And it is going to separate people fast. Not into "people who use AI" and "people who do not," because everyone uses AI now. It is going to separate the people who learned to hand off a job from the people who are still typing questions into a box.

That is it. One task. If it is good, you will figure out the rest on your own, I promise.