What Is ChatGPT Work? OpenAI's Task-Completing Agent, Explained

OpenAI shipped a lot on July 9. A new agent feature called ChatGPT Work, a desktop app reorganized around three pillars, a public beta for something called ChatGPT Sites, and, in the same breath, the death of its Atlas browser. Four announcements, one story: OpenAI wants ChatGPT to stop being a chat window and start being the thing that does your tasks.

This guide covers what ChatGPT Work actually is based on what OpenAI and early coverage have said, how it fits the rest of Thursday’s news, and how a normal office worker should approach it in week one.

What OpenAI actually announced

Start with the primary source. OpenAI’s release notes introduced ChatGPT Work on July 9, 2026, positioned for agent-based task completion. The same entry covers the desktop app being unified around Chat, Work, and Codex, and ChatGPT Sites entering public beta.

The release notes are, as usual, brief. For texture, iThinkDiff’s coverage describes ChatGPT Work as “an agent with built-in Codex that can complete tasks across web, mobile, and desktop.” That is their characterization rather than OpenAI’s official framing, but it matches the shape of the launch. The same report says Pro, Enterprise, and Edu accounts got access immediately, with Plus and Business following within days.

The distinction that matters for regular users is the one between answering and doing. ChatGPT has been able to tell you how to build a spreadsheet, draft a report, or reorganize a folder for years. An agent goes and does it, working through the steps itself and coming back when it is finished or stuck. Work is OpenAI’s dedicated home for that mode.

The Codex detail deserves a beat, because “Codex” sounds like a developers-only word. It is OpenAI’s coding agent, and if iThinkDiff’s description holds, having it built into Work means the agent can handle the technical half of ordinary tasks: manipulating files, wrangling data, automating a step that would otherwise need a script. You do not have to write code to benefit from an assistant that can.

The app reorganization tells you where this is going

On the same day, the desktop app was unified around Chat, Work, and Codex. That structure is a statement. Chat is conversation. Codex is coding. Work sits in the middle as the place where tasks get completed, and it gets equal billing with the chat interface that made the company famous.

And then there is Atlas. OpenAI announced on July 9 that it is shutting down its AI browser, with TechCrunch reporting the company’s conclusion that “the browser is a feature, not the destination.” Atlas was the bet that people would switch browsers to get an AI that could act on the web. Killing it on the same day Work launched is about as clear as strategy communication gets: the browsing and acting capabilities move inside ChatGPT, and the standalone browser goes away. If you adopted Atlas, plan your migration back to a regular browser and expect the agent side of your workflow to live in ChatGPT going forward.

ChatGPT Sites, the third piece, entered public beta the same day. OpenAI has said little about it in the release notes beyond the name and the beta status, so resist the urge to fill in the blanks. What matters for this article is the pattern: chat, agent, coding, and now some form of sites, all under one roof.

Who gets it, and when

Per iThinkDiff’s reporting, access at launch broke down by tier. Pro, Enterprise, and Edu: immediate. Plus and Business: within days of July 9, which as of this writing means it should be arriving on those accounts now if it has not already. Nothing has been said publicly about the free tier.

If you are on a work-managed Enterprise account, one wrinkle worth anticipating: your admin may control whether agent features are enabled at all. If Work is missing from your app and your colleagues on personal Plus accounts have it, the gap is probably policy, not rollout order. Ask IT before assuming a bug.

Sensible first tasks to try

The temptation with a new agent is to hand it something impressive. Do the opposite. Give it tasks where a failure costs you nothing and success teaches you the tool’s range.

A few that fit the bill for office work: have it take a folder of meeting notes and produce a single organized summary document. Ask it to turn a messy exported CSV into a clean, sorted spreadsheet with the columns you actually need. Give it a research task with a concrete deliverable, like comparing three software vendors against your written requirements and returning a table. Or hand it a repetitive formatting job you have been putting off, the kind of thing you know how to do and simply do not want to.

Notice what these have in common. Each one has a clear finish line, uses only material you provide, and produces an artifact you can inspect. That last part is the real value of early tasks: you learn how the agent works, where it gets confused, and how much checking its output needs before you trust it with anything that has your name on it.

Write the task like a delegation, not a wish. “Summarize my notes” invites a shapeless answer. “Combine the six files in this folder into one document, organized by project, with open action items pulled into a list at the top” gives the agent a spec it can either meet or visibly miss.

The part where you should be careful

An agent that completes tasks needs access to things. Files, at minimum. Possibly accounts, calendars, email, whatever a given task touches. This is the trade at the center of every agent product, and it deserves more thought than a reflexive click on an authorization prompt.

The sensible posture is narrow and revocable. Grant access to the specific folder a task needs, not your whole drive. When a task involving an account is done, revoke the connection rather than leaving it standing because you might use it again someday. Keep the agent entirely away from financial accounts, health portals, and password managers for now; no summary document is worth the blast radius if something goes wrong there.

Work accounts add another layer. Anything you connect on an employer-managed account potentially exposes company data to a system your company may not have vetted. If your organization has an AI policy, this is exactly what it exists for. If it does not, assume confidential material is off limits until someone with authority says otherwise. Boring advice. It stays boring right up until the day it is not.

Agents also fail differently than chatbots. A wrong chat answer sits there harmlessly until you read it. A wrong agent action has already happened: the file moved, the email sent, the wrong rows deleted. Early on, prefer tasks where the agent produces something new over tasks where it modifies something that exists, and keep copies of anything you let it touch.

Where this leaves you

If you are on Pro, Enterprise, or Edu, ChatGPT Work is already in your app and the best move is a week of small, checkable tasks on low-stakes material. Plus and Business users should see it within days of the July 9 launch if it has not landed already. Everyone else watches and waits.

The launch itself is easy to summarize. OpenAI folded its browser experiment, restructured its desktop app around doing rather than chatting, and gave the doing a name. Whether Work becomes the way ordinary people offload real tasks or joins the long list of AI features people tried twice depends on execution details nobody outside OpenAI knows yet. The release notes will say more as the rollout continues; check them before assuming this article’s tier breakdown still holds by the time you read it.