ChatGPT Work vs Claude Cowork: Two Desktop Agents Compared

If you are comparing ChatGPT agent with Claude Cowork, the first thing to know is that ChatGPT agent no longer exists. OpenAI’s help centre says it is no longer available and points to ChatGPT Work for longer, multi-step tasks. So the real comparison is ChatGPT Work vs Claude Cowork, and the two are closer than their names suggest. Both take an outcome rather than a prompt, plan the steps, work in the cloud while you do something else, reach your computer through a desktop app when a task needs local files or other apps, and come back with finished files. They differ in where the cloud and local halves meet, in how their browsers work, in what can start a task, and in which products sit around them.

Example tasks are covered elsewhere: twelve for ChatGPT in ChatGPT agent mode and what replaced it, and fifteen for Claude in Claude Cowork examples. This post compares the two products themselves.

What each one is

ChatGPT Work is one of the experiences inside ChatGPT, next to Chat and Codex. OpenAI’s guide to getting started with ChatGPT Work↗ (opens in a new tab) puts the split simply: use Chat for an answer, explanation or short draft, and Work when you want a task done with a clear outcome, such as a brief, a deck, an analysis, a recurring update or a file to review. You switch to Work above the message box, describe the result, and follow its progress. It shares its engine with Codex, and OpenAI says people who used Codex for non-coding work get the same core capabilities in Work.

Claude Cowork is Anthropic’s equivalent. Anthropic’s guide to getting started with Claude Cowork↗ (opens in a new tab) says it uses the same agentic architecture as Claude Code, with no terminal required: you describe an outcome, step away and come back to finished work. It is also changing shape. Anthropic has announced that Cowork and chat are becoming one Claude, rolling out gradually from Pro and Max, after which everything Cowork does is available from any conversation.

The lineage explains the old search term. OpenAI’s first browser agent was Operator; in 2025 it was folded into ChatGPT agent, and the Operator site is no longer accessible. ChatGPT agent has in turn given way to Work and its cloud browser. So “ChatGPT agent vs Operator” is now a history question, and the current answer to both is ChatGPT Work.

Where each one runs

• ChatGPT Work on the web and mobile runs in the cloud, on OpenAI-managed infrastructure, and keeps going when you leave the conversation. It cannot reach files or apps on your computer; you upload files, add them to a project, or connect an app.
• ChatGPT Work in the desktop app, on macOS, Windows and Linux, can run locally, with access to local files, apps and the browser. A control in the message box chooses Work locally or Cloud where both are offered; OpenAI suggests Cloud when the task should keep running after you close the app or turn off the computer.
• Claude Cowork runs its sessions in the cloud, in beta, on desktop, web and mobile, and they continue if you close your laptop.
• Claude Cowork needs the Claude Desktop app open whenever a task reaches your computer: local files, browser use and computer use.

The practical difference is small. In both, the cloud half is what lets work continue unattended, and the local half is what lets it touch a folder or a desktop app. ChatGPT makes you choose the location per task in the desktop app; Cowork runs in the cloud and calls on the desktop app when a step needs it.

What each one can reach

Files, other services and websites are the three things either agent works with, and each reaches them differently.

• Files. Work uses uploads, project files and your Library in the cloud, and local folders in the desktop app. Cowork uses files saved to your Claude account, and local folders through the desktop app.
• Other services. Work uses plugins, the installable bundles that carry apps (what ChatGPT used to call connectors) and skills, plus MCP servers your workspace allows. Cowork uses Claude’s connectors, skills and installed plugins, including custom connectors to any remote MCP server. In both, the connected account’s own permissions set the outer limit.
• The web. ChatGPT has a cloud browser that keeps its own cookies and sign-ins, a built-in browser in the desktop app, and a browser extension that lets it work in your own Chrome, Edge, Brave, Opera or Vivaldi tabs where you are already signed in. Cowork on desktop uses a browser built into the Claude Desktop app by default, and Claude in Chrome when it should work in your own browser.
• Other apps on your computer. Both can operate desktop apps through computer use, from their desktop apps, and both ask before each app. On the Claude side computer use is a beta on Pro and Max only; Team and Enterprise do not have it yet.

Sign-ins are where the two browsers differ most. ChatGPT’s cloud browser never sees your own browser’s passwords or sessions; when a site needs an account it pauses for a secure sign-in form. Using your own signed-in browser is a desktop-app and extension feature on both sides, and the reach it gives is the reason both vendors treat page content as untrusted. The permission model for Claude’s browser is in Claude in Chrome.

Approvals and checkpoints

Both let you choose how often the agent stops, and both keep a few stops you cannot switch off.

• ChatGPT Work on your computer. The desktop app has permission modes: Ask for approval, which OpenAI recommends for most work, Approve for me (automatic review) and Full access. Changing who reviews a request does not widen what it can reach.
• ChatGPT Work on websites. The cloud browser asks before visiting a new site by default, with Always ask, Auto approve and Always allow as settings, and asks again before anything consequential, such as a payment or a booking, even on an allowed site.
• ChatGPT apps. Each connected app asks for permission always, before making changes, or only before important changes, and workspace admins set the defaults.
• Claude Cowork. Anthropic’s page on using Cowork safely↗ (opens in a new tab) lists three modes: Manually approve, Automatically approve, where Claude screens each action for safety first, and Skip all approvals. Claude always asks before permanently deleting files, in any mode, and asks before using each app on your computer.

The advice from both vendors is the same: start on the most cautious setting, and stay there for anything that sends messages, spends money or touches sensitive accounts. Prompt injection, instructions hidden in a page, email or file, is the main risk for both, and both say to review what the agent opened when a task goes somewhere you did not expect.

Scheduling and triggers

Both repeat work. OpenAI’s page on scheduled tasks↗ (opens in a new tab) says ChatGPT can run tasks on a schedule, managed from Scheduled, and on eligible plans on the web and mobile it can also run a task when something happens in Gmail, Slack or GitHub: a new email from a sender, a message in a channel, activity on a pull request. Event triggers are not available in the desktop app. A desktop scheduled task that works with a local project needs the computer on and the app running.

Anthropic’s page on scheduling Cowork tasks↗ (opens in a new tab) offers hourly, daily, weekly, weekday and manual runs, created by describing the task to Claude or set up by hand. They run remotely, so they keep their cadence when your computer is asleep or the desktop app is closed; a scheduled task that needs local files or apps runs only locally. Time-based schedules are on both sides. Starting a task from an event in another app is, in the documentation today, a ChatGPT feature.

Plans, by name only

• ChatGPT Work: OpenAI lists it as included in Free, Go, Plus, Pro, Business, Edu and Enterprise, sharing usage with Codex. Individual features, such as the cloud browser, website sign-in and event triggers, depend on the plan, region and workspace settings.
• Claude Cowork: Pro, Max, Team and Enterprise, with Enterprise availability depending on the admin. Scheduled tasks are on all of those plans; computer use is on Pro and Max only.

Both draw on your plan’s usage. OpenAI says longer or more complex Work tasks may use more credits, and Anthropic says Cowork’s automatic approval mode uses more of your usage limit than the others, because of the extra checking. Check the current plan pages before choosing on cost.

Which one suits whom

• Already live in Google Workspace, Slack or GitHub and want work to start when something arrives: Work’s event triggers fit that directly.
• Want an agent that works inside the browser you already use, signed in to your tools: both do this from their desktop apps, through ChatGPT’s browser extension or Claude in Chrome.
• Your team’s projects, skills and connectors already live in Claude: Cowork uses them without a second setup, and the merge into ordinary conversations removes the mode switch.
• Your team’s projects and custom GPTs live in ChatGPT: Work is where they are going, as custom GPTs move to plugins.
• Mostly developers: both sit beside a coding agent from the same vendor, Codex for ChatGPT and Claude Code for Claude, sharing the same machinery.
• Regulated or sensitive work: compare the admin controls on the plan you would buy, not the consumer defaults. Both put plugins, connectors and approval modes under workspace policy.

Many teams will end up with both, because people choose the assistant that already holds their context. That makes the more useful question not which agent, but where the results of either one land.

Where the finished work goes

An unattended agent’s output ends up in one person’s task list or session history unless you give it somewhere shared to go. fenbs is a small task board that both can reach over MCP: ChatGPT as a custom app in developer mode where your plan allows it, and Claude as a custom connector. Each connects with a browser sign-in, holds your role on the board narrowed by the scopes you tick, and every task it adds, moves or comments on is recorded in History under its own name. A useful rule for either agent is to comment what it produced on the task it was working on and leave the move to Completed to a person.