How to Migrate a Custom GPT to a ChatGPT Plugin Before December 11 — and What Breaks When You Do

Short answer: migrate now, test everything, and know going in that the migration does not move your audience. OpenAI's retirement FAQ sets the date: 'Custom GPTs are scheduled to retire on Dec 11, 2026,' when 'custom GPTs and their GPT pages become inaccessible.' Migration is one click — My GPTs → Migrate to plugin — and 'the GPT's instructions become a skill within the new plugin,' knowledge files and connected apps come along. Four things don't: custom Actions, the selected model, your conversation starters and chats, and sharing. That last one is the trap. 'A migrated personal plugin starts private,' and on a personal, Plus, or Pro account there is no link, invite, or 'anyone with the link' to change that — the only documented exit is submitting it to the public directory for review. If you built a GPT for yourself, this guide is a 20-minute job. If you built one for students, clients, or the public, read the sharing section first.

The dates, from the live FAQ

Two useful details from the FAQ: 'you will still be able to migrate a GPT after retirement, although the GPT itself will no longer run,' so the deadline is for your users, not your data. And 'migration uses the GPT's latest published version. Drafts and unpublished edits do not transfer' — publish first.

What migrates and what doesn't

Why the output changes — and how to fix it

OpenAI says it plainly: 'A migrated plugin may respond differently.' Two mechanical reasons. First, the model is different — whatever you'd picked in the GPT builder is gone. Second, instructions become a skill, and skills are loaded differently from GPT instructions. Per OpenAI's developer docs, 'the model first sees skill metadata, including the name and description. It loads the complete instructions when the user's request matches the skill or the user invokes it directly.' And: 'ChatGPT may also use a relevant installed skill automatically when its description matches your request… the plugin will not necessarily run on every request.' A GPT applied your instructions to every message; a plugin applies them when it decides they're relevant.

That's the gap behind the complaints. Virtualization Review's David Ramel migrated his article-drafting GPT and 'its instructions survived. Required output didn't' — a sourced 'Evidence Pack' section vanished. Users on OpenAI's forum report instructions 'completely ignored' until they @-mention the plugin and tell it to follow the skill by name, output that 'changed headings, reorganized sections, omitted information,' and starters 'truncated during migration.' One found the skill 'does not seem to remain authoritative consistently across subsequent turns.'

• Invoke it explicitly. Type @ and pick the plugin, or open + → More. Don't rely on automatic selection for anything that matters.
• Make the skill description do work. The description is what the model reads first to decide whether to load your instructions; write it as 'Use for X, Y, Z' so matching is reliable.
• Put the format in the instructions, not in starters. Starters don't transfer one-to-one. If your GPT produced a fixed structure, state that structure explicitly in the skill text, and add an example output as a reference file.
• Run OpenAI's own test list: 'Compare familiar prompts and at least one harder case. Check that it: selects the right skill and follows your instructions. Uses the expected reference material. Produces complete answers or files in the required format. Has the tools and integrations the workflow needs.'
• Budget usage. Repairs happen in ChatGPT Work or Codex, which count against your plan; one user reported burning most of a five-hour and a weekly allowance fixing a single plugin. Edit through Plugin Creator ('@plugin-creator' plus the plugin's name) in short sessions.

Rebuilding Actions

If your GPT called an API, that connection is gone and nothing converts it automatically — OpenAI's docs contain no OpenAPI-to-MCP tool and no mention of custom actions beyond 'rebuild.' The FAQ's guidance: 'check whether an available app supports the tasks you need. Rebuilding the connection may require a custom MCP server and technical setup.' In practice, three paths, in order of effort:

• A supported app already does it. The plugin directory's connected apps (2,000-plus) cover most common services. If your Action hit Notion, Google Sheets, Slack, or a mainstream CRM, use the app's read or write actions instead.
• Build a small MCP server. OpenAI's guide: Build an MCP server. This is the equivalent of your old Action, speaking the open protocol every assistant now uses — our MCP explainer covers what that means. One MCP server per plugin.
• Host it on ChatGPT Sites. 'Hosting a plugin is available to all plans,' though in a workspace an admin must enable 'Upload plugins' and 'Create plugins with MCPs,' which are off by default on Enterprise.
• Then test: 'A rebuilt integration should not be assumed to provide every capability of the original action.'

Sharing: what's actually possible, by account type

Ghrist's screenshot of the migration dialog for his calculus tutor — 25,000-plus chats, shared with 'Everyone' — shows the warning in OpenAI's own words: 'Public sharing unavailable. Plugins can't be shared publicly. The plugin will stay private.' Here is what each kind of account can do, from the help center:

So what does a teacher with thousands of students, or a consultant with 25 customers, actually do? As of today there are two real options and one workaround. Option one: the public directory. Submission runs through an OpenAI API-platform organization, not your ChatGPT account: 'complete individual or business verification,' an 'Apps Management Write' role, and — if the plugin has an MCP connection — 'five positive test cases,' 'three negative test cases,' and 'a video walkthrough' for the review team. A skills-plus-files plugin qualifies in principle ('your plugin can include MCP connections, skills, or both'). No review time or fee is published. Once listed, anyone can find it, and whether Free accounts can install it 'depends on your plan, workspace, role, region.' Option two: a workspace. Put your users in a Business workspace and share there — practical for a 25-client consultancy, not for a lecture hall. The workaround users have found: a shared Project with the plugin attached, with the caveat one of them notes that 'all members of the project see conversations.' OpenAI hasn't endorsed it. OpenAI Support's written position, relayed by a creator on the forum, is that the absence of a sharing mechanism 'should NOT be interpreted as confirmation that this functionality will never exist.' That's a hope, not a plan — and December 11 is a date.

The migration checklist

• 1. Publish the latest version of the GPT. Drafts don't transfer.
• 2. Export what won't move: copy your conversation starters into a document; note the model you'd selected; list every Action and what it called; save example outputs in the format users expect.
• 3. My GPTs → Migrate to plugin. Read the three rows in the dialog — instructions and files, apps and actions, access and sharing — before confirming. The original becomes read-only.
• 4. Rebuild Actions as a supported app or an MCP server; test each one.
• 5. Fix the skill. Put format requirements in the instructions, add example outputs as reference files, write a description that matches the requests you want it to catch.
• 6. Test with @-mention and without. Familiar prompts, one hard case, the required format.
• 7. Sort out distribution. Personal account with an audience: start the directory submission now, since the review has no published timeline. Workspace: get an admin to enable Share plugins and Publish plugins to workspace, then share. Tell your users the old link dies December 11 and where the replacement lives.
• 8. Keep the GPT running until the deadline. It still works until retirement; your users don't have to switch before the plugin is right.

Our read

The migration is good engineering attached to a product decision OpenAI hasn't finished making. Instructions-as-skills and MCP-instead-of-Actions are the right architecture, and the one-click move is better than the nothing most platforms offer when they sunset a feature. But a GPT was two things — a tool and a URL you could hand to anyone — and the plugin replaces only the first. Ethan Mollick's observation landed because the dates check out: Plugins launched March 2023 and wound down April 2024, GPTs in November 2023 and the GPT Store in January 2024, Apps in October 2025, and now plugins again — a different thing under the same name, into which the apps were folded in July. The creators who got hurt each time were the ones who built an audience on the platform's distribution. Our advice is the same as our plugin-extensions explainer's: migrate because the tool is worth keeping, but if your GPT's value was its public link, treat the directory submission as urgent and keep a copy of your instructions somewhere you control. Our ChatGPT review notes the retirement where readers will look for it.