Scheduled tasks in ChatGPT Work
The feature that turns Work from a tool you operate into one that reports to you. It is also the fastest way to spend an agentic budget without noticing, so the cadence decision matters more than the prompt.
VERIFIED 2026-08-23 · OPENAI HELP CENTERFour run modes
| Mode | Fires | Fits |
|---|---|---|
| Once | A single future run | A deliverable due Monday that you brief on Friday |
| On a schedule | Repeating cadence | Weekly reports, month-end summaries |
| On a trigger | When an event happens | Work that should follow an event, not a clock |
| Monitoring | When something changes | Watching a source and reporting only on movement |
Source: OpenAI Help Center, ChatGPT Work and Codex.
Setting one up
- Get the task right manually first. Run it once by hand and accept the output. A scheduled task is a prompt you have stopped reading — it needs to already work.
- Put it in a project. A project keeps related chats, files and instructions together; starting Work from the project carries that context into every run, so results stay consistent instead of drifting.
- Choose the run mode — once, schedule, trigger, or monitor for changes.
- Decide what it may do unattended. You can review progress, answer questions, change direction and approve important actions while it runs — but only if you are there. Anything requiring approval will wait for you.
- Write the diary entry into the brief. Ask for a one-line summary of what changed since last run. Without it you are reading a full report to discover nothing happened.
The recurring cost trap
A daily monitor spends every day. Work draws from the same shared agentic usage and credit pool as Codex. A monitoring task that finds nothing still ran, still read its sources, and still cost you. Multiply by 30 before you pick "daily".
Three habits that keep it in hand:
- Set the slowest cadence that still answers the question. Most "daily" reports are weekly questions. Almost none are hourly ones.
- Prefer monitoring over polling. "Tell me when this changes" is cheaper than "check this every morning" for anything that changes rarely.
- Audit the list monthly. Recurring tasks are the line item people forget they authorised. The one you set up for a launch is still running six months later.
What we can't tell you yet
No verified per-plan cap on active scheduled tasks. Specific per-plan numbers appear in secondary coverage, but we have not confirmed them against OpenAI documentation and will not repeat them as fact. When OpenAI publishes them, this page changes and the badge date at the top moves with it.
Common questions
What can a scheduled task do?
A Work task can run once, repeat on a schedule or a trigger, or monitor for changes. While it runs you can review progress, answer its questions, change direction, and approve important actions.
Do scheduled runs cost the same as manual ones?
They draw from the same shared agentic usage and credit pool as any other Work or Codex task. A daily monitor is a task that spends every day, whether or not anything changed.
How many scheduled tasks can I have?
OpenAI has not published a per-plan cap that we can verify. Per-plan numbers circulate in secondary coverage; we are not repeating them as fact until they appear in official documentation.
Does a project help?
Yes. A project keeps related chats, files and instructions together, and starting Work from a project carries that context in — which makes a recurring task far more consistent run to run.