What are custom GPTs?
Custom GPTs are versions of ChatGPT that someone has set up for a specific job, with their own instructions, reference files, built-in tools and, optionally, actions that call outside APIs. You open one from the GPT directory or a shared link and chat with it like normal ChatGPT, except it already knows its role.
The most important fact in October 2026 is that custom GPTs are being retired. OpenAI's migration guide says "We're transitioning custom GPTs to plugins," and its retirement and migration FAQ schedules the retirement for December 11, 2026 across ChatGPT plans. If you rely on a GPT, plan the move now. This page covers how GPTs work today and what happens to them next.
Who can still build custom GPTs?
As of October 2026, per OpenAI's help articles GPTs in ChatGPT and Creating and editing GPTs:
| Account | Create new GPTs | Use existing GPTs |
|---|---|---|
| Free, Go, Plus, Pro (personal) | No | Yes, until retirement |
| Business, Edu | Where workspace settings and permissions allow | Yes, until retirement |
| Enterprise | Ended October 26, 2026 for affected workspaces (FedRAMP excluded) | Yes, until retirement |
OpenAI's troubleshooting guidance calls the personal-account limit an account eligibility restriction, not an error, so changing settings won't unlock it. If you built GPTs on a personal plan before the change, you can keep using them, and editing depends on your plan and permissions.
How is a custom GPT built?
Where creation is still available, the GPT editor offers two ways to work: a conversational builder you describe the GPT to, and a configuration view where you set each field yourself. The fields are:
- Name, description and conversation starters: what people see before they start, plus suggested first prompts.
- Instructions: the GPT's standing guidance, much like a system prompt.
- Knowledge: reference files the GPT can search.
- Model: the model selected for the GPT.
- Capabilities: built-in tools to turn on: web search, image generation, Canvas, and Code Interpreter & Data Analysis, plus Apps in workspaces that support them.
- Actions: connections to outside APIs.
Creating and editing GPTs works on the web; the mobile apps let you use GPTs but not build them. Test the GPT after every change, especially after uploading knowledge.
How should you write GPT instructions?
OpenAI's help center gives a few concrete tips: use an explicit step structure for multi-step workflows, prefer positive instructions ("do this") over long lists of prohibitions, and include examples when the GPT has to classify things. Keep rules in the instructions and reference material in knowledge files, not the other way around.
A compact pattern that follows that advice:
You help support agents draft replies for Acme Billing customers.
1. Read the customer's message and identify the issue type: billing, login, or bug.
2. Search the knowledge files for the matching help article.
3. Draft a reply under 120 words in plain language, citing the article title.
4. If no article matches, say so and suggest escalating to the billing team.
Example: "I was charged twice" is a billing issue.
Instructions are the part of a GPT most worth keeping. They are also exactly what OpenAI's migration turns into a skill, so writing them as clear, numbered procedures pays off twice.
How do knowledge files work?
Knowledge is a set of files the GPT can draw on when answering. Per OpenAI's help center, as of October 2026:
- up to 20 files per GPT,
- up to 512MB per file,
- up to 2 million tokens per text or document file.
OpenAI recommends using knowledge for reference material rather than rules and keeping files text-forward. Clean Markdown, plain text or well-structured documents work better than scanned PDFs. Split a large manual into topic files with clear headings, so the right passage is easier to find.
What are GPT actions?
Actions let a GPT call outside services. OpenAI's GPT Actions docs describe them as a way for "ChatGPT users to interact with external applications via RESTful APIs calls outside of ChatGPT." Under the hood they use function calling: you describe the API with a schema, and the model picks the right endpoint and fills in the parameters, the same idea explained in what is tool calling.
- Authentication: the editor offers None, API Key or OAuth, per OpenAI's authentication docs.
- Privacy policy: OpenAI's troubleshooting guidance says public actions need a valid privacy policy URL.
- Apps or actions, not both: OpenAI's GPTs and sharing docs say "GPT builders can use either connected apps or custom actions, but not both in the same GPT."
- Workspace domains: admins can limit actions to approved domains, and "If no domains are allowed, custom GPT actions cannot execute."
Actions are the part that doesn't survive migration. OpenAI says custom actions do not transfer to plugins and must be rebuilt with a supported app or a custom MCP server.
How do sharing and the GPT Store work?
OpenAI's sharing and publishing article describes sharing levels that include anyone with the link and the GPT Store, and which ones you get depends on your account, workspace settings and permissions.
Publishing to everyone can require a completed builder profile, and only a verified domain can be used for the website link on it. As of October 2026, personal accounts can't publish new GPTs to the GPT Store, and a GPT that uses Apps can be blocked from public link sharing and GPT Store publishing.
On privacy, the help center says builders can't see the individual chats people have with their GPT. A GPT also starts fresh: it doesn't use your saved memories, custom instructions or past conversations.
What replaces custom GPTs?
Plugins. OpenAI describes a plugin as "a reusable package for a workflow that can include skills, apps, or both." Per OpenAI's migration guide and retirement FAQ:
| GPT part | After migration to a plugin |
|---|---|
| Instructions | Become a skill |
| Knowledge files | Copied as reference files |
| Connected apps | Carried over |
| Custom actions | Not transferred; rebuild with an app or MCP server |
| Selected model | Not carried over |
| Conversation starters | Not transferred one to one; save useful prompts |
| Sharing | New plugin starts private; reshare it yourself |
Migration uses the latest published version of the GPT, so drafts and unpublished edits are left behind. In workspaces, creators can choose Migrate to plugin from My GPTs, and once migrated the original GPT becomes read-only. OpenAI says "At retirement, custom GPTs stop running and leave the GPT directory." Migration options can differ by account, so follow the notice in your own ChatGPT account.
Save your GPT instructions before December
Whatever you do with the plugin migration, copy your GPT's instructions and conversation starters out now. They are plain text, and they are the part you will want again in a plugin, a ChatGPT project or another assistant. Kipt keeps prompts and Agent Skills as Markdown in a private GitHub repo you own, with Cmd+K search, so a GPT's instructions stay usable after the GPT itself is gone. If they read like a procedure, the free SKILL.md generator turns them into a SKILL.md in your browser. For background on the format, see Agent Skills.