ChatGPT Work — the replacement for Agent Mode
ChatGPT agent mode is gone. OpenAI's help center now opens its agent article with: "ChatGPT agent is no longer available. Use ChatGPT Work for longer, multi-step tasks and finished deliverables." This page covers what Work is, where it runs, what it can touch, and how it differs from the agent mode many older guides (including an earlier version of this one) describe.
ChatGPT now has three experiences: Chat for quick conversational help, Work for longer multi-step work that ends in a finished deliverable, and Codex (desktop app) for software. Work is available on web and mobile for eligible paid plans, and in the desktop app, where it can also use local files and apps with your permission. For browser tasks that agent mode used to handle, OpenAI points to the cloud browser.
What Work is for
OpenAI's own split (ChatGPT Work and Codex):
- Chat: ask questions, search, brainstorm, get quick help.
- Work: research a topic, analyse information, or create a document, spreadsheet, presentation, report or Site.
- Codex: write or debug code, run tests and commands, review changes, work with a repository. Desktop app only.
Where it runs
- Web and mobile: Work runs in the cloud. Pick Work from the dropdown (mobile) or the toggle (web).
- Desktop app: choose ChatGPT in the top-left menu, then Work from the toggle at the top. Work here can open a local folder and use local files and desktop apps with your permission. Messages and task context may still be stored in the cloud even when work runs locally.
- Sync: cloud Work chats sync across web, mobile and desktop. Local chats stay on the computer that ran them.
- Projects: from a Project, choose Work to start a Work chat that uses the project's files and instructions.
Models
- GPT-6 Sol and GPT-6 Luna are Work and Codex models. They are not available in regular Chat conversations.
- GPT-6 Astra is available in Work and Codex. Plus plans and Business Standard seats include limited Astra usage. Pro $100, Pro $200 and Business Premium seats can use their full allowance on it. Astra uses the allowance faster than GPT-5.6 Sol.
- The model picker offers Default, which combines models and reasoning levels on one slider, or a specific model with its reasoning levels. What you see depends on your plan and workspace.
Work uses the same usage structure as Codex: an included allowance, then optional credits. OpenAI's Codex pricing page explains the allowance, but its examples are coding tasks, so Work usage will vary.
Run it on a schedule or a trigger
Work tasks can run once, repeat on a schedule, monitor for changes, or fire on an event (Scheduled tasks):
- Event-triggered tasks run in Work when a supported event happens in a connected app: new Gmail messages, new Slack channel messages, or GitHub pull-request activity. Available on Plus, Pro, Business, Enterprise and Edu, not on Free or Go. Create or edit triggers on the web or mobile; the desktop app can show them but not edit their conditions. For Slack triggers, add @ChatGPT to every channel the task watches.
- Active scheduled-task limits: 3 on Free and Go, 5 on Plus, 10 on Business and Edu, 15 on Pro and Enterprise. Free can schedule at most once a day; paid plans can run hourly.
- Manage everything on the Scheduled page.
Voice
ChatGPT Voice adds a live voice interface to Work on web and mobile, and to Work and Codex in the desktop app (macOS and Windows). On web and mobile, actions that need approval must be approved on screen, because spoken approval isn't supported. Voice in Work uses your daily Voice minutes. Only one Voice conversation can run at a time.
What changed from agent mode
| Agent mode (retired) | ChatGPT Work | |
|---|---|---|
| How you started it | Tools menu or /agent in the composer | Choose Work from the toggle or dropdown |
| Limits | Monthly message caps (e.g. 40/month on Plus, 400/month on Pro) | Shared Work & Codex allowance, then credits |
| Browser | Built-in virtual browser (Operator merged in) | Cloud browser for supported browser workflows |
| Your files | Only files uploaded to the chat | Desktop app can use local folders and apps with permission |
| Triggers | Could repeat a finished task on a schedule | Schedules plus Gmail, Slack and GitHub event triggers |
The agent-mode column is from OpenAI's archived agent article, kept for readers following older tutorials.
For workspace admins
- Manage Work per role in Workspace settings → Permissions & roles. Work Cloud allows cloud tasks across web, mobile and desktop. Work Local alone allows desktop-local work but no cloud tasks. Codex Local is controlled separately.
- Set starting defaults for Work & Codex (model, reasoning level, speed, Fast mode) in Workspace settings → Models. Chat keeps its own default.
- Enterprise, Edu and ChatGPT for Healthcare admins must turn on Allow event-triggered scheduled tasks before members can create them.
When to use Work, Chat, a self-hosted agent, or the API
| Job | Best fit |
|---|---|
| Quick question, brainstorm, rewrite | Chat |
| Research brief, spreadsheet, deck or report from several sources | Work |
| Watch an inbox, channel or repo and act on new items | Work with an event-triggered task |
| Programmatic use inside your own product | OpenAI API |
| Always-on automation on your own hardware or with any model | Hermes or OpenClaw |
| Coding inside a repository | Codex, Kilo Code or Claude Code |
Safety habits that still apply
- Grant Work only the folders and connected apps a task needs. It can act on what it can reach.
- Anything an agent reads (web pages, emails, documents) can contain instructions aimed at it. Treat that content as data. See our security guide.
- Keep an eye on approval prompts, especially for tasks that send, publish or delete.
← Back to ChatGPT hub · Next: Memory
Official sources
Checked against OpenAI’s help center and chatgpt.com/pricing (US prices) on 2026-09-27. If anything here disagrees with these pages, trust the vendor docs.
← Back to ChatGPT hub · Next: Memory feature →