Now that OpenAI has launched ChatGPT Work, its version of Claude Cowork, ChatGPT has three modes: Chat, Work and Codex.
You already know what Chat does. The real question is what Work adds and when Codex is still the better choice.
TL;DR
Is Work using a better model?
Not really.
Chat, Work and Codex can use the same GPT-5.6 model family. Work and Codex have separate model and reasoning controls, but Work is not simply a smarter version of Chat.
The default Power setting currently uses GPT-5.6 Sol with medium reasoning. You can increase the reasoning effort or use Ultra on eligible plans to split a complex task across subagents. See the current model options.
What Work changes
Chat can already search the web, read files and use skills or plugins.
Work can take that context and run a longer task without waiting for you to prompt every step.
Think of it like this:
Files, projects and memory provide the context.
Skills define your process.
Plugins connect your tools.
Schedules decide when it runs.
Work runs the full task.
For example:
A skill can define the format. Plugins provide access to Slack and Drive. Work runs the steps.
Can Work run loops?
Yes.
For a loop within one task, give Work a clear definition of done and ask it to check its output before finishing.
For a loop that repeats over time, use a scheduled task. Scheduled tasks can run later, repeat on a cadence or monitor for a change. They can also use skills and plugins. Learn about scheduled tasks.
Web or desktop?
| Version | What it can access |
|---|---|
| Web | Uploaded files, connected tools and approved websites |
| Desktop, local | Files, apps and the browser on your computer |
| Desktop, cloud | Hosted tasks that continue after you close the app |
Web tasks cannot work directly inside a local folder. Local desktop tasks can, but your computer and the app need to remain available.
What makes Codex different?
Codex keeps the technical work visible: repository instructions, diffs, shell commands, tests, worktrees and pull requests.
Work can run code. Codex can create documents. The main difference is how you review the result.
Work returns finished output. Codex shows you the implementation.
Use Work for a report, spreadsheet, deck, Site or workflow. Use Codex when the result needs to be tested, versioned and committed.
The same task in both
| Task | Work | Codex |
|---|---|---|
| Website | Research, copy or a standalone Site | Changes inside the website repository |
| Data | One-off analysis or report | Reusable scripts or a data pipeline |
| Automation | A workflow across plugins and apps | A task that checks or changes a repository |
How this maps to Claude
ChatGPT Work is closest to Claude Cowork. Codex is closest to Claude Code.
OpenAI can also import supported skills, plugins, projects and recent work from Claude Code or Cowork. See the import options.
If the work is spread across your tools, use Work. If it lives in a repository, use Codex.