ChatGPT Work is an agent built into ChatGPT that OpenAI launched on July 9, 2026. Instead of answering a question or drafting a paragraph, you hand it an outcome (a competitor comparison, a status report, a cleaned-up spreadsheet) and it pulls context from your connected apps, breaks the job into steps, and works on its own for hours before handing back a finished document, spreadsheet, slide deck, or website. It runs on OpenAI’s GPT-6 model family: GPT-6 Astra is the flagship model, with GPT-6.1 Sol, GPT-6 Sol, and GPT-6 Luna available only inside Work and Codex, not in regular ChatGPT conversations, and draws down your ChatGPT plan’s usage allowance the same way Codex does, rather than costing anything extra on its own.
Starter · Level 1 of 3
What ChatGPT Work actually is, in plain terms
Think of ChatGPT Work less as a smarter chatbot and more as a delegate sitting on top of the chatbot you already know. OpenAI folded the agent technology from Codex, its coding tool, into ChatGPT itself, and describes Work as software that “gathers information across your apps and workflows to create finished materials” and “stays with complex projects for hours by breaking them into smaller steps and completing them independently,” according to OpenAI’s launch announcement. The new ChatGPT desktop app now presents three modes side by side: Chat for quick back-and-forth, Work for longer research and deliverables, and Codex for software development, per OpenAI’s Help Center.
The practical difference from a normal ChatGPT conversation comes down to five things.
| Regular Chat | ChatGPT Work | |
|---|---|---|
| What you give it | A prompt, refined message by message | An outcome, described once |
| Where it gets context | Whatever you paste or attach | Connected apps and files, pulled automatically or @-mentioned |
| How long it runs | Seconds per reply | Up to several hours on one task |
| What comes back | Chat text | A finished document, spreadsheet, slide deck, or website |
| Your role | Drive every step | Approve the plan, answer check-ins, review the result |
What it can do across connected apps today
The launch directory covers more than 1,400 connectable apps and services: Slack, Microsoft Teams, Google Drive, SharePoint, Gmail, Outlook, calendars, CRMs, and project trackers among them. ChatGPT decides which connected app is relevant to a request automatically, or you can type "@" followed by an app name to pull its context in directly. Before anything runs, Plan mode lays out the steps it intends to take so you can correct course early rather than discover a wrong turn hours in.
- Documents and analysis. Sheets, slide decks, written reports, and now Sites, a public-beta feature that turns a request into an interactive web app or dashboard shareable by URL, which ChatGPT can keep updated as the underlying information changes.
- Scheduled and event-triggered Tasks. A job can run once, repeat on a schedule, or sit as an ongoing monitor, reviewing new Slack messages every week to refresh a meeting agenda, for example, or turning a stream of customer feedback into a running list of prioritized ideas. OpenAI has since added event-triggered (webhook-based) tasks for Plus, Pro, Business, Enterprise, Edu, and ChatGPT for Healthcare plans, so Work can also start automatically on a supported connected-app event, such as a new Gmail message, a new Slack channel message, or GitHub pull request activity, instead of waiting for a schedule or a manual nudge.
- Desktop-only extras. On the Mac and Windows desktop app, Work can also reach local files and installed apps with your permission, use a built-in browser, and take over your computer directly through Computer Use, clicking, typing, and moving files for a one-time job or a recurring one. Web and mobile Work runs entirely in the cloud and cannot touch files stored on your computer, per OpenAI’s Help Center.
Access is rolling out in stages rather than all at once. Pro, Enterprise, and Edu accounts got Work on web and mobile starting July 9; Plus and Business accounts followed over the next several days. The desktop app, which bundles Chat, Work, and Codex in one place, is available on every plan, including Free, though Free accounts only get it through that desktop app rather than on web or mobile. As part of the same release, OpenAI folded the agentic-browsing work from its standalone Atlas browser into ChatGPT and began sunsetting Atlas as a separate product.
Pricing has also moved since launch. Pro access is no longer a single tier: it now comes as Pro $100, Pro $200, and Pro $500, and the $500 tier, along with eligible Enterprise and Edu workspaces, also gets a lower-latency Ultrafast mode for Work and Codex. If you’re comparing against an older figure you saw for a flat Pro plan, check which tier it actually referred to before assuming it still applies.
Where it still needs a human check
Evidence check. This explainer is built from OpenAI’s July 9, 2026 launch post and OpenAI Help Center articles on ChatGPT Work, most recently rechecked October 3, 2026 — desk research, not hands-on use of the product. OpenAI has already moved Work and Codex from the original GPT-5.6 model to the GPT-6 generation and restructured Pro pricing and connected-app automation since launch, so treat plan-by-plan availability, model names, and connector details here as accurate as of this writing, not fixed. Scheduled for a freshness recheck by 2026-11-03.
A few things are worth sitting with before you hand it anything that matters.
- You are the approver, and that is by design, not a limitation to work around. OpenAI’s own materials say you decide what ChatGPT Work can access, when it checks in with you, and when it needs your approval before acting. An "auto-review" system also uses OpenAI’s more advanced models to check higher-stakes actions on connected tools and APIs before they happen. Both are safeguards worth keeping switched on, not friction to disable once you trust the tool.
- Cost is metered, not flat. There is no published per-task price. Usage draws down your plan’s included allowance at a rate that varies with how much work a task actually requires, the same structure Codex already uses. A single ambitious multi-hour run can use up a meaningful share of a month’s allowance, so the first few tasks are worth treating as a measurement exercise, not a habit yet.
- The product keeps changing underneath its own name. OpenAI has already swapped the underlying model generation once since the July 2026 launch and restructured Pro-tier pricing along the way, and connector and automation behavior for any single app can still change without much notice. Confirm your plan’s current tier, model access, and connected-app permissions inside your own account before planning a workflow around a specific feature.
Who it’s actually useful for vs who should skip it
ChatGPT Work fits best when you already know a workflow cold and just want the assembly work off your plate: a monthly performance deck built from the same few sources every time, a competitor benchmark you’d normally block out a week for, or a lead-triage pass across a CRM, an inbox, and call notes. It also fits if your account already sits on Pro, Enterprise, Edu, Business, or Plus, since those are the plans with web and mobile access as of launch, and if the apps that hold your real context (Slack, Drive, a CRM, a calendar) are ones you’re comfortable connecting.
It fits less well, or not yet, if any of the following is true. You’re on the Free plan and need it on web or mobile: Free gets the desktop app, but cloud Work still requires a paid plan. Your task is a single quick question with no real multi-app research behind it: that’s still just a Chat, and routing it through Work wastes a slice of your metered allowance for no benefit. You want something that quietly watches an inbox or a CRM all day and reacts the instant something changes. Work is task-based: you start it, it runs for a bounded stretch, and then it stops, which is a different shape from an always-on monitoring tool. Or nobody on your side is actually going to read the Plan-mode outline and check the result before it reaches a client or a stakeholder. In that case the risk isn’t the tool, it’s skipping the one step that makes it safe to use.
How to try it safely the first time
A worked first run beats reading about one. Here’s a concrete example you can adapt: turning a week of Slack activity in one project channel into a one-page status update.
- Open ChatGPT and switch modes. On desktop, click "ChatGPT Work" in the mode switcher at the top left. On mobile, use the Chat/Work dropdown at the top of the screen. On web, select Work from the mode menu. Regular Chat stays one click away if you want it — desktop keeps a "Quick chat" button in the sidebar for exactly that.
- Connect only the one app the task needs (Slack, in this example) instead of turning on every available plugin at once. Work can only reach what you’ve connected and enabled, so narrower access for a first run is also a safer one.
- Describe the outcome, not a chat prompt: something like "Pull the last seven days of activity from the #launch-prep Slack channel and draft a one-page status update with sections for wins, blockers, and next steps."
- Read the Plan mode outline before approving it. Check that it picked the right channel and date range, and adjust anything that looks off before letting it start. This is the cheapest correction you’ll make in the whole run.
- Leave check-ins and action approvals turned on for this first attempt, even though it means a few extra clicks. That setting is what lets both OpenAI’s auto-review and your own judgment catch a wrong turn before it happens instead of after.
- Treat what comes back as a draft, not a finished deliverable. Skim names, dates, and numbers you already know before the document goes anywhere near a teammate or a client.
- Only after that supervised run should you consider turning the same request into a recurring Scheduled Task.
If you already lean on ChatGPT for everyday planning, Work is the next step up from asking it to turn a messy to-do list into a weekly plan or to organize your ongoing projects in one place — same assistant, a longer leash. Start with one task you already know well, keep approvals switched on, and judge the result against how long that task actually takes you by hand before you build anything bigger around it.