ChatGPT scheduled tasks: what they do, and where the ceiling is

What it is

ChatGPT can run one-off or recurring tasks on a schedule, and on paid plans it can also react to events rather than clocks. You describe what you want in ordinary language — "every weekday at 7am, summarise anything new about my industry" — and it confirms the schedule back to you. Results arrive in the chat, with notifications if you enable them.

What you actually get, by plan

Checked September 2026. These change — verify in the app before planning around a number.

Event-triggered tasks are the more interesting half: on paid plans they can fire from Gmail, Slack or GitHub activity rather than a clock. That turns "every morning" into "when this happens", which is a genuinely different capability.

Read that table as a ceiling, not a price

The cost of a ChatGPT scheduled task is bundled into your subscription, so it feels free at the margin. The real constraint isn't money — it's the number in the second column.

Fifteen active tasks is the most any tier offers. If your ambition is "watch sixty competitor pages" or "check this every five minutes", the platform doesn't become expensive, it becomes unavailable. There is no tier that sells you a thousand scheduled prompts, because that isn't what the product is.

Setting one up

Just ask, in the same conversation you'd ask anything else:

Every weekday at 7am, check the RSS feeds for [sources] and give me
the three items most relevant to [your topic], one line each.
Skip anything you've already sent me.

Manage what exists from the Scheduled section — pause, edit, delete. Three prompting habits make a real difference:

Fix the output shape. "Three bullets, one line each, no preamble" stops you receiving an essay on the day it feels chatty.

Say what to skip. Without it, a daily digest will cheerfully repeat yesterday's news. Scheduled tasks don't share memory across runs the way you'd expect, so the instruction has to be explicit — and even then it's not a substitute for real state.

Ask for silence. "If nothing meaningful has changed, reply exactly: nothing today." Otherwise it will always find something, because you asked it to.

Where it falls down

The task limit. Covered above, and it's the one that ends most ambitious plans.

No reliable memory between runs. Each run largely starts fresh. Anything that depends on "what did I see last time" — change detection, deduplication, a running total — is the one thing this shape is worst at, and it happens to be what most monitoring actually needs.

Tasks in projects can't reach project files. A documented and surprising limitation: a scheduled task created inside a project doesn't get access to that project's uploaded files.

Not available in voice chats or GPTs.

Non-deterministic. Same input, different phrasing, occasionally different judgment. Fine for a digest you read. Not fine for anything downstream that parses the output.

It lives in an account. Tied to one login, one subscription, one company's roadmap. If that's carrying something your business depends on, it's worth noticing.

What it's genuinely good for

Personal routines where judgment is the job and volume is low: a morning briefing, a weekly summary of a topic, a nudge when something in your inbox needs attention, practice questions, a Friday retrospective prompt. All of these need a model every time — they're not rules dressed up as intelligence — and all of them run once a day.

That's the sweet spot, and inside it this is the least-effort automation available anywhere.

When to move it

Three signals: you've hit the task limit and want more; you need it to reliably remember what it saw last time; or it's become something you'd be annoyed to lose. Any of those and it's worth an hour porting it to a scheduled script, where you get unlimited tasks, real state, and a cost of zero.

The overview has the full comparison and the arithmetic on when that's worth doing.