ChatGPT automation in 2026 has two lanes. ChatGPT alone covers scheduled tasks, event-triggered tasks on Gmail, Slack and GitHub, plugins and the Work agent, within limits of 3 to 15 active tasks by plan. A workflow platform or the API covers everything else: other triggers, higher volume, and writes to your systems. Custom GPTs are not part of either lane any more, because OpenAI retires them on December 11, 2026.
ChatGPT automation in 2026 comes in two lanes: what ChatGPT can run by itself (scheduled tasks, event-triggered tasks, plugins and the Work agent) and what still needs a workflow platform such as n8n or Zapier, or the OpenAI API. The first lane is included in your plan and capped at 3 to 15 active tasks; the second has no such cap but has to be built and maintained.
Limits, dates and prices checked October 2026 against OpenAI's help center articles on scheduled tasks, plugins, ChatGPT Work, workspace agents, the custom GPT retirement FAQ and ChatGPT Plus, plus the ChatGPT pricing page and the API deprecations page.
One change reshapes every older guide, including the earlier version of this page: OpenAI is retiring custom GPTs. If a tutorial tells you to automate something by building a GPT, skip it. The failure notes below are the limits OpenAI documents, not results from our own test runs. For what the API lane costs per run, start with our LLM API pricing comparison.
ChatGPT workflow automation: the two lanes
Use ChatGPT alone for personal, recurring work where a message or a document is the result. Use a platform when the workflow is a business process: other systems, other people and volume.
- 01Is it one prompt on a timetable?
Scheduled task.
- 02Does Gmail, Slack or GitHub start it?
Event-triggered task.
- 03Is it a procedure you reuse?
Plugin skill.
- 04Does a team share it?
Workspace agent.
- 05Other triggers, volume or writes?
n8n, Zapier or the API.
| ChatGPT alone | Workflow platform or API | |
|---|---|---|
| How often it can run | Up to once an hour on paid plans; once a day on Free | Any interval, including every minute |
| Triggers | A schedule, or Gmail, Slack and GitHub events | Webhooks, forms, databases and hundreds of apps |
| Volume | 3 to 15 active tasks; 30 event runs an hour | Limited by your plan or server |
| Actions on other systems | Through connected apps; approval can pause the task | Any API call, with your own guardrails |
| Output | A message, notification or document | Structured data written wherever you need it |
| When it fails | The task pauses; you check Scheduled | Error workflows, retries and execution logs |
| Cost | Included in the ChatGPT plan | Platform plan or server, plus API tokens |
The left column is cheaper and takes minutes to set up. The right column is where you go when you hit one of the left column's limits, and the limits are specific enough to plan around.
What ChatGPT can automate by itself in 2026
Four features do the work. Three are open to individuals; the fourth is for teams.
- Scheduled tasks. One-time or recurring prompts, plus monitoring tasks that check for a change and notify you when it happens. Available on every plan, with results by push notification or email.
- Event-triggered tasks. Tasks that fire on a new Gmail message, a new message in a Slack channel or pull request activity on GitHub. They run in Work and need Plus or above.
- Plugins. Packages that combine skills (instructions and workflow guidance) with connected apps. They work in Chat, Work and Codex, and they are what OpenAI now recommends in place of custom GPTs.
- Workspace agents. For Business and Enterprise workspaces: agents built for a repeatable task, shared with teammates, usable in Slack, run on a schedule or triggered through an API.
Underneath those sits Work, which OpenAI describes as an agent for longer, multi-step work and finished deliverables, next to Chat for quick conversational help. Its usage draws on your plan's allowance, so a long job costs more of it than a short one.
Event-triggered tasks are also capped at 30 runs an hour and 720 a day across all of them. Source: OpenAI Help Center, Scheduled tasks in ChatGPT, checked October 2026.
Five active tasks on Plus is the number that decides most people's setup. It is plenty for a briefing, two monitors and a couple of inbox rules. It is not a business automation layer.
Custom GPTs are retiring: what replaces them
Custom GPTs are scheduled to retire on December 11, 2026, and the change covers every ChatGPT plan. For two years "build a GPT" was the standard answer to "how do I automate this in ChatGPT", so this is the part of your setup to check first.
- Sep 17, 2026GPT-to-plugin migration target
OpenAI's target date for the Migrate to plugin flow. Not every account gets it at once.
- Oct 23, 2026gpt-4 and gpt-3.5-turbo leave the API
API workflows that still name them stop working.
- Nov 30, 2026Agent Builder shuts down
OpenAI points builders to its Agents SDK or to ChatGPT workspace agents. ChatKit remains available.
- Dec 11, 2026Custom GPTs retire
Existing GPTs keep working until this date.
- Feb 11, 2027Enterprise deferral ends
Only for Enterprise workspaces with an approved deferral.
Source: OpenAI Help Center and API deprecations page, checked October 2026.
If you created a GPT you rely on, OpenAI's FAQ says to open My GPTs and select Migrate to plugin when the option appears for your account. Three details in that FAQ matter for anything automated:
- Custom actions do not transfer automatically. A GPT that calls an external API needs that integration rebuilt, so allow extra time.
- The original becomes read-only after migration. Finish your edits first.
- Scheduled tasks never supported GPTs. The tasks article lists GPTs and voice chats as unsupported, so a GPT was never a way to run something on a timer.
Ten ChatGPT automations, with the limit that breaks each one
The first seven run inside ChatGPT. The last three look similar but belong on a platform, and the final column says why. Each note is a documented limit, so you can check it against your own case before you build.
| Workflow | Built with | Plan | Failure note |
|---|---|---|---|
| 1. Morning briefing | Scheduled task, daily | Free and up | Free runs once a day in a loose window. Exact times and hourly runs need a paid plan |
| 2. Change monitor (a page, a price, a topic) | Monitoring task | Free and up | Counts against your active task limit until its end condition is met |
| 3. Gmail triage and reply drafts | Event-triggered task on new mail | Plus and up | Filters only by sender or subject. An action that needs approval pauses the task |
| 4. Slack channel digest | Event-triggered task on channel messages | Plus and up | ChatGPT has to be added to every channel it watches |
| 5. Pull request summary | Event-triggered task on GitHub activity | Plus and up | Only authorized github.com repositories and supported pull request events |
| 6. Weekly report as a finished document | Recurring scheduled task | Plus and up | A task created inside a project cannot read that project's files |
| 7. A saved procedure (brief, review, rewrite) | Plugin skill | Plugin directory is on all plans | Do not build it as a custom GPT: they retire December 11, 2026 |
| 8. High-volume email or ticket classification | n8n or the API | API billing | ChatGPT caps event-triggered runs at 30 an hour and 720 a day |
| 9. Form or webhook into a CRM or sheet | n8n or Zapier | Platform plan plus API | Tasks start only on a schedule or on Gmail, Slack and GitHub events; validate fields before writing |
| 10. Multi-step pipeline with approvals | n8n or Zapier | Platform plan plus API | Needs branching, retries and logs that a single task does not give you |
Workflows 3 to 5 are the new ground this year. Before event-triggered tasks, reacting to an email meant a platform. Now a Plus account can do it, as long as a sender or subject filter is enough and you are happy to approve anything that sends or changes data.
Workflows 8 to 10 are where people lose a weekend trying to bend ChatGPT into a job it is not shaped for. If the work arrives by form or webhook, runs hundreds of times a day or has to land as clean fields in a database, go straight to a platform. Our GPT and n8n use cases guide has an importable workflow for exactly that pattern, and the n8n OpenAI integration guide covers the credential and the nodes.
Copy these task prompts
These are starting instructions for workflows 1 to 6. Each one says what to do when there is nothing to report and stops short of sending anything, which are the two habits that keep a task useful after the first week. Replace the parts in square brackets.
1. Morning briefing (scheduled, weekdays 8am)
Search for news from the last 24 hours about [topic]. Give me at most five
items, each with one sentence on why it matters to [my business] and the
source link. If nothing meaningful happened, reply "Nothing new today".
2. Change monitor (monitoring task)
Check [URL] once a day. Tell me only when the price drops below [amount]
or the page shows it is back in stock. Stop after you have told me once.
3. Gmail triage (event-triggered, sender filter: [client domain])
When a new email arrives, summarise it in two sentences, list any dates or
deadlines it mentions, and draft a reply in my usual tone. Do not send it.
4. Slack digest (event-triggered, channel: #feedback)
When a message is posted, label it bug, request or praise and add one line
to the running summary for the day. Mention me only for bugs that block a
customer.
5. Pull request summary (event-triggered, repo: [owner/repo])
When a pull request is opened, explain what changed in plain English, list
the files touched, and flag anything that affects sign-in, payments or
database migrations.
6. Weekly report (scheduled, Fridays 4pm)
Write a one-page report with three sections: what shipped, what slipped,
and what needs a decision. Use only the sources I have connected, and mark
any figure you could not confirm.Two of them lean on rules from the tasks article. Number 2 uses an end condition, because a monitoring task can stop itself once the condition is met and free up a slot. Numbers 3 to 5 never send, because an action that needs approval pauses the task until you review it; a draft waiting for you does not.
For teams: workspace agents
Workspace agents are the shared version of all this, for ChatGPT Business and Enterprise. You build one from a template or by describing the job, review the draft plan, preview it with a sample prompt, then publish it to teammates or the whole workspace. An agent can have tools, apps, custom MCP servers, skills and files attached, can be used from Slack, and can run on a schedule or be triggered through an API.
Two cautions from OpenAI's article. On Enterprise, workspace agents are off by default and an admin has to enable them. And access follows role-based permissions, so people only see and run the agents they have been given. If your team's process currently lives in a shared custom GPT, this is the closest replacement to evaluate before December.
How to automate ChatGPT tasks: your first one
Start with one scheduled task and get it right before you add a second. The steps follow OpenAI's tasks article.
- 1Write the prompt as a finished instruction
Say what to check, what to include and how long the answer should be. The task runs the same words every time.
- 2Put the schedule in the same message
For example: every weekday at 8am, summarise new posts on these three sites. ChatGPT confirms the task in the chat.
- 3Turn on notifications
Settings, then Notifications. Choose Push, Email or both.
- 4Find it in Scheduled
The Scheduled page lists each task with its schedule and results, and lets you pause, edit or delete it.
- 5Read the first three runs
Edit the instruction from the task panel until the output is what you would have written yourself.
- 6Keep the conversation
Deleting the chat a task belongs to pauses the task.
For an event-triggered task the path is slightly different: connect Gmail, Slack or GitHub under Settings, then Apps; open Work and describe the event and the action; then review the Trigger, Condition and Prompt it proposes. Read the condition carefully. A Gmail rule that matches every incoming message will spend your hourly cap on newsletters.
When to move a workflow off ChatGPT
Move it when the limit you keep hitting is structural, not a matter of wording the prompt better.
- You are the only person who reads the result
- A schedule, Gmail, Slack or GitHub starts it
- Once an hour is often enough
- A message or a document is the output
- You are fine approving anything that sends
- A customer or teammate depends on it
- A form, webhook or database row starts it
- It runs more than 30 times an hour
- The output is fields written to another system
- It needs retries, logs and an alert when it fails
Moving does not mean starting over. The prompt you refined in a task becomes the instruction in a model node, and the judgement you built about what good output looks like becomes your test set. What changes is the bill, which goes from a flat subscription to tokens per run. Our ChatGPT API workflows guide walks through the API side, and AI agent automation covers the cases where a fixed workflow is not enough.
What each lane costs
Inside ChatGPT the cost is the subscription. Scheduled tasks are included on every plan, Plus is $20 a month, and Pro comes in $100, $200 and $500 tiers with higher limits (checked October 2026). Work usage counts against the plan's allowance, and OpenAI sells credits for more.
On a platform you pay twice: for the platform or the server it runs on, and for API tokens. The token side is usually the smaller number. At OpenAI's list prices, classifying a 700-token email with an 80-token answer on gpt-6-luna costs about $0.00011, so a thousand runs cost 11 cents. Long writing jobs on a larger model cost cents each. Those figures are arithmetic on list prices with example token counts, not measurements.
Before you rely on any ChatGPT automation
- The instruction says what to do when there is nothing new to report
- Nothing sends, pays or deletes without a person approving it
- You have counted your active tasks against your plan's limit
- No step depends on a custom GPT or on Agent Builder
- Sensitive details are not in a task title or instructions you plan to share by link
- Someone looks at the Scheduled page each week for paused tasks
- API workflows name a current model ID, not gpt-4 or gpt-3.5-turbo
- You timed the manual version first, so you know the automation is worth keeping
The fifth check comes from the sharing rules: a shared task link shows its title and full instructions to anyone who can open it. The sixth matters because a task can pause for inactivity or a pending approval, or stop responding when a connected app loses its authorization, and the Scheduled page is where you find out.
Sometimes the workflow you keep rebuilding is one other people would pay for. At that point it is a product. Our AI SaaS Builder program has no ChatGPT, n8n or Zapier lessons; it covers building that product in code, from validating the idea to a Claude API feature, a Supabase database, deployment and Stripe billing.
ChatGPT automation: FAQ
What is ChatGPT automation?
It is getting ChatGPT to do recurring work without you typing the prompt each time. In 2026 that means scheduled tasks that run on a timetable, event-triggered tasks that react to Gmail, Slack or GitHub activity, plugins that package reusable instructions with connected apps, and the Work agent for longer jobs. Anything beyond those limits moves to a workflow platform or the OpenAI API.
Can ChatGPT run tasks automatically on a schedule?
Yes. Scheduled tasks are available on every ChatGPT plan, including Free. Free users can schedule a one-time task or a recurring task no more than once a day, in flexible windows such as morning or night. Paid plans support runs up to once an hour and exact delivery times. Results arrive by push notification, email or both. Checked October 2026.
How many scheduled tasks can I have in ChatGPT?
OpenAI's help center lists the active task limits as 3 on Free and Go, 5 on Plus, 10 on Business and Edu, and 15 on Pro and Enterprise, checked October 2026. If you hit the limit, pause or delete a task, or wait for one to finish. Event-triggered tasks have a separate cap of 30 runs an hour and 720 a day across all of them.
Are custom GPTs being discontinued?
Yes. OpenAI's help center says custom GPTs are scheduled to retire on December 11, 2026, on all ChatGPT plans, with February 11, 2027 for Enterprise workspaces that have an approved deferral. Existing GPTs keep working until then. OpenAI recommends migrating them to plugins from My GPTs, and warns that custom actions do not transfer automatically.
Can ChatGPT react to new emails or Slack messages?
Yes, with event-triggered tasks. They respond to a new Gmail message, filtered by sender or subject, to new messages in a Slack channel where ChatGPT has been added, or to pull request activity on GitHub. They run in Work and need a Plus, Pro, Business, Enterprise or Edu plan. Free and Go accounts cannot create them.
When do I need n8n or Zapier instead of ChatGPT?
Move to a workflow platform when the job needs a trigger ChatGPT does not offer, such as a form or a webhook from your own app, runs more often than once an hour, exceeds 30 events an hour, has to write to other systems without pausing for approval, or needs fixed branching, retries and logs. ChatGPT tasks suit personal recurring work; platforms suit business processes.
How much does ChatGPT automation cost?
Scheduled tasks are included in ChatGPT plans, so the cost is the subscription: Free with limits, or Plus at $20 a month, checked October 2026. Work usage draws on your plan's allowance. If you move a workflow to the API, you pay per token instead, which is cents per thousand runs for classification on a small model and more for long writing jobs.
Free · no card needed
Start with the free Starter Kit.
A unit economics calculator from AI SaaS Builder, plus four more working files.
Outgrown scheduled tasks? Build the thing properly.
All Access includes AI SaaS Builder, which takes a workflow worth selling and turns it into a product: idea validation, a Claude API feature with structured output, Supabase, deployment and Stripe billing. It has no ChatGPT or n8n lessons. The other three programs, live coaching and the private community are included.
Not sure which lane your workflow belongs in?
Describe it in the free Discord and ask. It is also the place to compare notes on moving a custom GPT to a plugin before the December deadline.