Claude Cowork might finally have a competitor. On July 9, OpenAI launched ChatGPT Work, an upgrade to its desktop app that used to be called Codex and is now simply called ChatGPT. Like Claude Cowork, you give ChatGPT Work an outcome. It gathers what it needs from your apps and files, breaks the work into steps, and works for as long as it takes (sometimes for hours) to hand back finished material: the actual spreadsheet, the slides, the document, even a working web page. This is a plain-English guide to what ChatGPT Work is, what it can and can't do, and how it compares to Claude Cowork, so you can decide which one belongs in front of your team.
What ChatGPT Work actually is
OpenAI describes ChatGPT Work as "an agent in ChatGPT that helps you take on more ambitious tasks." In plain terms, four things happen when you use it.
How ChatGPT Work turns an outcome into finished material
Instead of "write me an email," you say something like "look at last quarter's support tickets, find the five most common complaints, and build me a one-page summary with a chart." It reaches into the apps and files you've connected to pull in what it needs, maps out an approach, and stays on the project until it hands back a spreadsheet, a slide deck, a document, or an interactive web page you can share by link. The engine underneath is GPT-5.6, OpenAI's newest model, which is only available on the paid Pro account and went live the same day.
The real shift: from advisor to doer
For the last three years, "using AI" meant chatting with AI. You'd ask a question, copy the answer, and finish the task yourself. It was helpful, but most people didn't see huge efficiency gains, because you were still the one doing the work. ChatGPT Work, just like Claude Cowork, moves the AI from advisor to doer. You're no longer typing a question and getting words back. You're asking for an outcome and getting a working deliverable, saved to your drive.
It can sound like a small thing. It isn't. Once you start working this way, you may stop using normal chat almost entirely.
Plan first, then let it work
The first reaction to "it works on its own for hours" is usually, "what is it doing while it does that?" The honest answer is that your judgment stays in the loop the whole time. That happens through a few mechanisms.
- Plan mode. Before it starts, ChatGPT Work gathers context, asks you questions, and lays out a step-by-step plan. Spend the time to get this right, even if it takes longer than you'd like. Don't just skim and approve. The goal is to leave the AI as few judgment calls as possible.
- Check-ins and approvals. You decide what it's allowed to touch, when it should pause and ask, and which actions need your explicit sign-off. When it hits something that needs a human call, it stops and asks rather than guessing.
- Steering mid-job. You can watch its progress, answer its questions, and redirect it partway through if it's heading the wrong way. The catch: you have to be able to tell when it is heading the wrong way. If you can't, find someone who can.
Think of these tools as assistants who do the legwork and check with you at decision points, not robots you set loose on your systems. That flow isn't friction to switch off once you trust it. It's the mechanism that keeps your judgment in charge while the tool does the volume. Worth noting: this is exactly how Claude Cowork works too. None of it is unique to ChatGPT Work.
Connected apps, scheduled tasks, and Sites
ChatGPT Work is only as useful as what it can see. OpenAI's answer is a directory of more than 1,400 "apps" it calls Plugins (its word for connectors to outside tools). You bring a tool's context into a task by typing @ and the app's name, and ChatGPT can suggest relevant ones as it works. The more of your tools you connect (email, calendar, CRM, document storage), the more of your real work it can pick up and run with, instead of you playing courier between apps.
It can also run tasks on a schedule, not just when you're sitting there: once, on a recurring schedule, when a specific event happens, or continuously as a monitor. OpenAI's own examples include reviewing your Slack updates every week to refresh a meeting agenda, turning a running stream of customer feedback into a prioritized list of product ideas, and updating a deck automatically whenever new feedback arrives by email. This is the same "assign it and walk away" capability Claude Cowork has. Start small, and improve over time.
Beyond spreadsheets, slides, and docs, ChatGPT Work can also produce interactive web pages and small web apps, a feature OpenAI calls Sites (currently in public beta): live dashboards, a project tracker, a launch calendar, an internal reference page, each shareable by a link, and kept up to date as the underlying information changes. For most business users, that's a genuine leap. You can describe a dashboard and get one your team can actually open and use.
ChatGPT Work vs. Claude Cowork
The core promise is nearly identical. Both take an outcome instead of a prompt, work across your files and connected apps, run multi-step for extended stretches, return finished deliverables, run scheduled unattended jobs, keep you in control with plans and approvals, and run across web, mobile, and desktop. So the decision comes down to the details, and the details matter, especially if you're rolling this out to a team.
- 1,400+ app directory and a built-in browser (absorbing the Atlas experiment)
- Bets on Sites, its shareable web apps, as the output format
- More connectors, but not always better ones (no native Make or n8n plugin, for example)
- Skills install manually, or by coaxing ChatGPT to install them
- Still feels designed for developers
- Models widely felt to be smarter and easier across many tasks, not just code
- Widely regarded as the stronger long-form writer
- Fewer connectors, but generally the ones you actually want
- Skills and connectors add in one click, right in the interface
- Built for everyone, with Claude Code as the dev-focused mode when you need it
Same promise; the fit is where they part ways