In 2026 the two big AI labs shipped the same idea within six months of each other: an agent that does your actual work — files in, finished deliverables out — instead of chatting about it. Anthropic launched Claude Cowork in January ("Claude Code for the rest of your work"). OpenAI answered in July with ChatGPT Work, folding its Codex agent into ChatGPT for everyone. If you're deciding which one to learn, this is the comparison we wish existed.
Where each one came from
Both products are descendants of coding agents, and it shows — in a good way. Claude Cowork is the Claude Code engine (the terminal agent programmers have used since 2025) pointed at folders of documents instead of repositories. ChatGPT Work is what happened after OpenAI merged Codex into ChatGPT: a "Work" toggle that hands your goal to an agent on a cloud computer. The lineage matters because both inherited the thing that makes coding agents trustworthy: explicit plans, permission gates, and reviewable output.
The head-to-head
| Claude Cowork | ChatGPT Work | |
|---|---|---|
| Launched | January 2026 (research preview) | July 2026 |
| Included with | Claude Max, Team, Enterprise | ChatGPT Plus, Pro, Business, Edu, Enterprise (usage-metered) |
| Platforms | macOS + Windows desktop, web, iPhone/iPad/Android (beta) | Web, desktop (macOS/Windows), iOS — one merged app with Chat and Codex |
| Core model | Point at a folder; agent works with real files, pauses for decisions | Give a goal; agent plans, runs on a cloud computer, returns an artifact |
| Deliverables | Documents, spreadsheets, decks, contract redlines, research reports | Docs, spreadsheets, decks, PDFs, images (GPT Image 2), interactive Sites |
| Integrations | Connectors (Gmail, Calendar, Drive, Slack…) + plugin marketplace | 1,400+ plugins with bundled skills, @-mentions |
| Scheduling | Server-side scheduled tasks (cloud sessions) | One-time, recurring, event-based, and monitoring triggers (cloud only) |
| Computer / browser use | Browser tasks + computer-use research preview | Agent browser, Chrome extension, computer use with its own macOS cursor |
| Phone story | Steer running sessions from mobile; cloud state syncs | Remote tab: voice master-thread that spawns and steers desktop sessions |
| Reusable instructions | Skills — follow your account across surfaces (incl. Excel add-in) | Skills — but cloud and local skills don't sync across the divide |
Where Cowork wins
- File-first work: point it at a messy folder and the workspace model feels native — inputs, outputs, and the brief all live together.
- Spreadsheet depth: the Claude Excel add-in builds models with live formulas, scenario dropdowns, and self-verifying checks tabs.
- Calmer trust model: folder access is opt-in per session and the approval loop is consistent everywhere — easiest story to roll out carefully.
- Skills portability: one skill works in Cowork, chat, and Excel without cloud/local forking.
Where ChatGPT Work wins
- Artifact breadth: native PPTX/DOCX/PDF plus images and hosted interactive Sites — chat to shareable web app in minutes.
- Plugin surface: 1,400+ integrations with bundled skills, and the agent picks the right one without being told.
- Automation triggers: event-based ("when her reply arrives") and monitoring modes, not just clock schedules.
- Voice and Remote: triaging your inbox and spawning work sessions by voice on a walk is a genuinely new workflow.
The gotchas nobody mentions
- ChatGPT Work is usage-metered — a 30-minute agent run consumes real allowance, so measure a task before scheduling it weekly.
- Work's local desktop sessions don't sync and their scheduled tasks silently skip if the laptop is closed — always schedule from cloud surfaces.
- Cowork's mobile apps are still labeled beta and roll out plan-by-plan, Max first.
- Both ship work you'll be judged on. The review habit — read it before you forward it — is the actual skill, and neither product can do it for you.
So which should you learn?
If your team runs on Claude (Max/Team/Enterprise), learn Cowork. If you're on ChatGPT Plus/Pro/Business, learn Work. If you have both or neither, pick by workload: heavy spreadsheets and document review lean Cowork; decks, sites, images, and inbox automation lean Work. And genuinely — don't agonize. Task shaping, plan review, approval gates, skills, and scheduling are the same five skills in both products, and they're the same five skills the next agent will use too.
What to do next
Pick one, run one real workflow through it this week — an expense report, a weekly update, a research brief — with plan mode and approvals on. Measure it against how long the task takes you by hand. That single measured pilot will teach you more than any comparison table, including this one.