What Is ChatGPT Work? OpenAI’s Autonomous Work Agent, Explained

OpenAI launched ChatGPT Work on July 9, 2026, and the pitch is simple enough to sound almost suspicious: describe a job, and the software goes off and does the whole thing. Not a draft paragraph, not a suggestion in a sidebar, but a complete assignment carried out across the apps and files you already use, returned to you as a finished result. It is OpenAI’s clearest step yet from "chatbot that answers questions" toward "agent that gets work done," and it is worth understanding what that actually means before you decide whether it belongs anywhere near your job.

What ChatGPT Work is

ChatGPT Work is an autonomous AI agent. You hand it an assignment, it gathers the context it needs from your connected applications, files, and workflows, and then it produces a finished deliverable rather than a set of instructions for you to follow. The outputs are the ordinary artifacts of knowledge work: documents, spreadsheets, presentations, reports, and even websites.

The distinction that matters is between assistance and delegation. A standard chatbot waits for each prompt and answers in a conversation. An agent like this one takes a goal, breaks it into steps on its own, reaches into the tools where your information lives, and works through the task without you supervising every move. If you want the broader category, we cover it in our explainer on AI agents, and ChatGPT Work is one of the more ambitious consumer-facing examples of that idea to date.

How you actually use it

The interaction model is built around the assumption that you will not be sitting and watching. You can start a task from your phone, walk away, and let it run autonomously in the background. When you want to know where things stand, you check progress from the web or the desktop app.

That flow changes the rhythm of using the tool. Instead of a back-and-forth session that holds your attention, ChatGPT Work is closer to handing an assignment to a contractor: brief it, let it go, review the result. In practice that means the quality of your initial brief carries more weight than it does with a chatbot, because there are fewer natural moments to correct course midway.

• Kick off from anywhere. Tasks can be started on mobile and continue running once you close the app.
• Runs unattended. The agent works in the background rather than requiring a live session.
• Check in on your terms. Progress is visible through the web interface or the desktop app whenever you want to look.

Standing automations, the headline feature

The feature OpenAI is leading with is standing automations. You set one up once, and it runs on a schedule without being asked again. The example the company uses is a Slack digest: the agent scans your Slack messages on a set cadence and generates a summary, so you get the roundup automatically instead of prompting for it each time.

This is the part that pushes ChatGPT Work past being a faster way to do one-off tasks. A recurring automation turns the agent into something closer to background infrastructure, quietly handling routine jobs on a timer. It is also, not coincidentally, the part that raises the most questions about oversight, because an automation that runs on its own is an automation you are not watching each time it fires. More on that below.

The engine underneath

ChatGPT Work is powered by GPT-5.6, OpenAI’s current flagship model. That matters because agentic behavior leans heavily on the model’s ability to plan across multiple steps, hold context from several sources at once, and decide which tool to reach for at each stage. A weaker model tends to lose the thread on long, multi-part jobs; a stronger one can hold a plan together well enough to finish it.

It is worth separating the model from the product, though. GPT-5.6 is the reasoning engine; ChatGPT Work is the scaffolding around it that connects to your apps, schedules automations, and manages tasks that run over time. The model makes the agent capable, but the product design is what decides how much reach and autonomy that capability is given. This is a different surface from the real-time, conversational direction of ChatGPT Live; the two reflect two distinct bets OpenAI is making about how people will use its models.

Where it sits in the market

ChatGPT Work is OpenAI’s push into workplace automation, and it lands in a category that is suddenly crowded. The most direct point of comparison is Anthropic’s Claude Cowork, which chases the same idea of an agent that carries out full assignments rather than answering questions. We will publish a proper head-to-head on that separately, so this is not the place to referee it; the useful context is simply that the two largest AI labs now both believe the next competitive frontier is autonomous work, not conversation.

That framing tells you something about where the industry thinks this is heading. For a couple of years the pitch was "an assistant that helps you write and think." The pitch now is "an agent that does the task while you do something else." Whether that pitch survives contact with real workflows is the open question, and it is one you should hold onto rather than take on faith.

The honest limits

ChatGPT Work is new, and newness is a caveat in its own right. Autonomous agents that act across your applications and files raise real questions that a conversational chatbot does not, and those questions are worth sitting with before you wire one into your working life.

• Trust and accuracy. An agent that produces a finished report or spreadsheet can be confidently wrong, and a polished deliverable can hide its mistakes better than a rough draft does. Results should be reviewed, not shipped on sight.
• Permissions and reach. Connecting an agent to your apps and files means granting it access to that data. The convenience and the exposure are the same act, so it is worth being deliberate about what you connect.
• Oversight. Standing automations run without you in the loop, which is the point and also the risk. A scheduled job that quietly drifts off course can do so repeatedly before anyone notices.

None of this means the tool is unsafe or unusable. It means the responsible way to adopt an agent is gradually: start with low-stakes tasks, keep a human reviewing the output, and expand its remit only as it earns your trust. Delegation to software follows the same logic as delegation to people, and the early days of any new hire call for checking the work.

Frequently Asked Questions