The short answer
OpenAI retires custom GPTs on 11 December 2026 on every ChatGPT plan, with Enterprise workspaces that have an approved deferral given until 11 February 2027. The replacement is a ChatGPT plugin, a bundle of skills, apps and MCP servers. To keep an existing GPT, open My GPTs and choose Migrate to plugin: its instructions become a skill and its knowledge files become reference files, but custom actions do not transfer automatically. For a new assistant, write it as a skill.
What changed, in plain English
A custom GPT was a saved version of ChatGPT with your instructions, a few knowledge files and optional actions that called other services. OpenAI is replacing them with plugins. A plugin bundles skills (reusable instructions and workflows), apps (connections to outside services) and MCP servers. Skills are listed on every ChatGPT plan on the pricing page, and plugins work across ChatGPT and Codex.
The key dates, from OpenAI's custom GPT retirement and migration FAQ (see also our ChatGPT guide):
- 11 December 2026: custom GPTs retire on every plan.
- 11 February 2027: the later date for Enterprise workspaces with an approved deferral.
- Existing conversations stay with the original GPT and do not move into the plugin, so copy anything you still need before the retirement date.
The same thing is happening at Google. For personal accounts of adults, Gemini's Gems become read-only on 17 November 2026 and are removed on 17 December 2026, with skills as the replacement. Workspace business accounts move in March 2027. The idea everyone is converging on is the same: a skill, a folder with a SKILL.md file holding the instructions plus any reference files.
What you need first
- A list of the GPTs you rely on. For each, note whether it uses knowledge files, custom actions (calls to outside services) or a specific model.
- A copy of each GPT's full instructions in a document you control. You will want them whichever route you take.
- The published version up to date. Migration uses the latest published version, so drafts and unpublished edits do not transfer. Publish anything you want to keep.
- Three or four real test tasks per GPT, so you can check the replacement behaves the same.
How to migrate an existing custom GPT, step by step
- In ChatGPT, open My GPTs.
- On the GPT you want to keep, choose Migrate to plugin. The option appears once migration reaches your account.
- Review the details and follow the prompts.
- Check what came across: the instructions become a skill, knowledge files become reference files and connected apps come across as apps. Sharing does not carry over: a migrated personal plugin starts private, so share it again with the people who used the GPT.
- Rebuild what did not. The model you selected does not carry over, and custom actions do not transfer automatically; they may need a custom MCP server, which is a job for a developer.
- Run your test tasks. A skill is loaded when ChatGPT judges it relevant, so if it does not kick in when expected, make the skill's description more specific about when to use it.
If you only use a GPT someone else built, you cannot migrate it. Check its notice for a replacement or ask the creator.
How to build a new assistant the 2026 way
The durable approach is to write your assistant once as a skill, then load it wherever you work. Start by writing the instructions in a document with three parts:
- Name and description. A short name and one or two sentences on exactly when the skill should be used ("Use when drafting replies to customer complaints about delivery"). The description is what tells the AI to load it.
- Instructions. The role, the steps to follow, the format of the output and the rules it must never break.
- Reference files. Your style guide, price list or policy, as separate files.
Then put it where your team works:
- ChatGPT. OpenAI's plugin documentation describes a Plugin Creator you talk to as @plugin-creator, and a plugin's skills live in a skills folder, each with its own SKILL.md. The documented process includes editing manifest files, so treat it as a developer task for now. Workspace admins can publish a plugin from Personal in ChatGPT Plugins and choose which roles get it. For a simpler home for instructions and files, ChatGPT's projects are listed on Plus.
- Claude. A Claude Project is the closest thing to a custom GPT that works today on every plan, including Free (up to five projects). Click Projects, then + New Project, add files with the + button in the knowledge panel and click Set project instructions. Our guide to Claude Projects walks through it.
- Gemini. Open the sidebar, select Customize, then Skills, then Create and Create manually. Paste the name, description and instructions and click Create. Call it by typing / and its name; up to 100 skills can be active at once, with 100 MB of files per skill.
A worked example: moving a "proposal writer" GPT
Say your sales team uses a GPT that drafts proposals from call notes, with your price list and two past proposals as knowledge files and an action that pulls client details from the CRM. First, copy its instructions into a document and publish any pending edits. Migrate it from My GPTs; the instructions arrive as a skill and the three files as reference files. The CRM action does not come across, so for now the team pastes client details into the chat and you log the action as a job for IT. Run four real call notes through the plugin and compare against the old GPT's output. It skipped the pricing table once, so you add "always include the pricing table from the price list" to the skill and tighten the description to "use when writing a sales proposal from call notes". Keep the instructions document: the same text makes a Claude Project or a Gemini skill in five minutes if part of the team works there.
Common mistakes and how to fix them
- Waiting until December. Fix: migrate and test now, while the old GPT still runs for comparison.
- Losing drafts. Only the latest published version migrates. Fix: publish first.
- Assuming actions still work. Fix: list every action and plan its replacement, usually a connector or MCP server.
- A vague skill description. The skill never loads. Fix: say exactly when to use it, with the words people actually type.
- Putting secrets in instructions. Fix: keep passwords, keys and confidential client data out of skills and reference files.
Which tool should you use for custom assistants now?
If your team lives in ChatGPT, migrate to plugins and keep your instructions in a SKILL.md you control. If you want the simplest setup with files and instructions today, Claude Projects is the easiest. If you are in Google Workspace, Gemini skills replace Gems. Compare the assistants in ChatGPT vs Claude and ChatGPT vs Gemini, read what a custom GPT is, or see ChatGPT alternatives.
Prompts to copy
Turn GPT instructions into a skill
Here are the instructions from my custom GPT: [paste]. Rewrite them as a skill with: a short name in lowercase with hyphens, a description of one or two sentences saying exactly when to use it, then numbered instructions, the output format and the rules it must never break.
You get: A clean SKILL.md draft you can use in ChatGPT, Claude or Gemini
Audit your GPTs before the deadline
Here is a list of our custom GPTs with what each does and what it uses: [list]. For each, say what will migrate (instructions, knowledge files, apps), what will not (model choice, custom actions) and what we need to rebuild, then rank them by risk.
You get: A migration plan ordered by what could break
Sharper skill description
This skill is not loading when people need it. Description: [paste]. Typical requests that should trigger it: [examples]. Rewrite the description so it clearly matches those requests and does not match [requests that should not trigger it].
You get: A description that makes the skill load at the right time
Before and after test
Compare these two answers to the same task, one from the old GPT and one from the new plugin: [paste both]. List every difference in content, format and tone, and say which instruction to add to the skill to close each gap.
You get: The exact edits that make the replacement match the original
Reference file from a messy document
Turn this document into a clean reference file for an AI assistant: [paste]. Keep only facts the assistant needs, use headings and short bullet points, and flag anything that looks out of date.
You get: A reference file the assistant can rely on