Custom GPTs Retire December 11: What to Move To and How
OpenAI is replacing custom GPTs with plugins. Instructions and files move with one button, custom actions do not, and some GPTs should go somewhere else entirely. Here is how to decide and do it before the deadline.
Your custom GPT stops working on December 11
If your team has a GPT that writes quotes, answers staff questions or turns call notes into follow-up emails, it has about five weeks left as of early November. OpenAI's help center says custom GPTs are scheduled to retire on December 11, 2026, and that on the retirement date the GPTs and their pages become inaccessible. FedRAMP workspaces and Enterprise workspaces with an approved deferral get until February 11, 2027. Everyone else should plan on December.
OpenAI's replacement is a plugin, and if your GPT is just instructions and a few files, moving it is a short job. The trouble is custom actions, the API connections some GPTs use to read or write data in another system. Those don't move at all. This guide covers what transfers, how to do the move and when a plugin is the wrong destination.
What to move to
For most people the answer is a plugin, since that is the only path with a built-in migration button. A plugin bundles skills (reusable instructions) with connected apps, and ChatGPT can pick it up automatically when a request matches, or you can call it with an @ mention. If the GPT is a helper that a person opens and asks things, this is your default.
Three other destinations are better in specific cases.
If the GPT should run without someone starting it, look at ChatGPT Work scheduled tasks, or workspace agents if you are on a Business or Enterprise plan. Workspace agents can be shared, placed in a Slack channel, run on a schedule or triggered by an API call. That suits a job like a weekly report much better than a chat helper does.
If the GPT faces customers, a plugin is the wrong shape. Plugins live inside ChatGPT, and the people using yours need a ChatGPT account, access to the plugin and a plan where plugins work. Migration doesn't carry over public sharing settings either. A website chatbot on your own domain is the usual replacement. One build on this site shows a customer service bot with knowledge base retrieval, lead capture and booking, sold at $3,500 (builtwithagents.ai/strategy/live-build-3500-customer-service-chatbot). Before you pay for one, read this site's honest take on whether a chatbot is worth it (builtwithagents.ai/blog/is-ai-chatbot-worth-it-small-business-website).
If you would rather leave ChatGPT, Claude is the cleanest alternative for instruction-driven work. Anthropic's pricing page lists skills on every plan, including Free, and one builder documented reusable Claude skills for reviewing construction contracts and writing scopes of work (builtwithagents.ai/strategy/claude-skills-construction-contract-review-scopes). That listing is a method walkthrough and has no measured results. Don't move to Gemini Gems, though. Google says personal Gems become read-only and convert to skills on November 17, 2026, with business Workspace accounts following in March 2027. Gemini skills are the thing to build, and Google says most default Gem tools, including Canvas and Deep research, don't work with skills yet.
A worked example: the roofing company's quote GPT
This one is an illustration with assumed details, so you can see how the move plays out.
A 12-person roofing company has a GPT called Quote Writer. Its office manager built it last year. It has about 900 words of instructions, three PDFs (price sheet, warranty terms, sample proposals), Google Drive connected as an app, and one custom action that pulls job details from the company's CRM. Four estimators use it each week.
On Monday, the office manager publishes the last edits and saves five real requests as test prompts, such as a 24-square architectural shingle tear-off on a 7/12 pitch. One hard case goes on the list too: an insurance claim for a partial replacement, where the warranty language changes. Then comes the click on Migrate to plugin.
The instructions become a skill, the PDFs become reference files and Drive comes across as an app. The CRM action does not. So the plugin writes solid quotes but can't pull a job record by itself. The plugin directory has no CRM app that covers the record they need, so two things happen. Estimators paste job details into the prompt for now, and a developer is booked to rebuild the connection as a custom MCP server. Testing takes an afternoon. It shows that the partial-replacement case skipped the warranty paragraph, so the office manager tightens one instruction and retests.
By Friday the plugin is shared with the workspace, and the estimators install it. The new bill is zero, since the seats were already paid for. The developer's time is the only real cost.
How to migrate your GPT, step by step
Step 1: List every GPT you depend on
Open My GPTs in each account or workspace you use, and write down each GPT's name, who created it, who uses it and whether it has custom actions. The account matters: OpenAI says the timeline and permissions of the account where a GPT was created are what apply. If you only use someone else's GPT, you can't migrate it yourself. Check its notice, then ask the creator what the replacement will be.
Step 2: Publish your final edits and save a copy
Migration uses the latest published version, so drafts and unpublished edits are left behind. Publishing doesn't mean sharing publicly. After migration the original becomes read-only and you can't delete it, so finish editing first. OpenAI doesn't ask for this, but paste the instructions into a document and keep the knowledge files in a folder. It takes two minutes and gives you a copy no matter what happens.
Step 3: Save your test prompts
Pick three to five prompts you actually use, plus one hard case that has tripped it up before. Run them now and save the answers. OpenAI warns that a migrated plugin may respond differently, and without these you will be guessing whether it got worse.
Step 4: Run the migration
Go to My GPTs and select Migrate to plugin. If you don't see it, migration may not have reached your account yet, plugin access may be off in your workspace, or you are signed in to the wrong account. In a managed workspace, ask your admin to check that plugins are enabled. Don't leave this to the last week, since OpenAI says a target date doesn't guarantee every account can migrate on that day.
Step 5: Pick a model and test against your saved answers
The GPT's selected model doesn't carry over, so check which one you are on. Run your saved prompts and compare. OpenAI's checklist is a good one: does it select the right skill, use the expected reference files, produce complete answers in the required format and have the tools it needs? If answers drift, fix the skill's instructions rather than your prompts.
Step 6: Rebuild custom actions
Check the plugin directory for an app that does what your action did, because that is far easier than building one. If there is none, the fallback is a custom MCP server, which is a technical job and depends on what your account allows. OpenAI also warns not to assume a rebuilt integration can do everything the original could, so test each read and write separately. Anything that writes to a live system should keep its approval step on.
Step 7: Share it and move people over
The plugin starts private, and nobody gets access automatically. On Business and Enterprise you can share it with your workspace, and managed workspaces can prefill the original GPT's audience in Share Plugin. A public release needs a separate submission process. Tell people the plugin's name, have one person install and test it, and then set a date to stop using the old GPT. Their old conversations stay accessible after retirement, so nobody needs to export chats.
What this costs
OpenAI's FAQ doesn't list a migration fee. As of November 2026, its pricing page shows you can use plugins on every plan, but creating and sharing them starts with Go at $8 a month, then Plus at $20. Business is $20 per seat billed annually, or $25 monthly, and that is where workspace sharing comes in. If you are on Free and rely on a GPT you built, check your in-product notice for your options.
For Enterprise workspaces, OpenAI says using a migrated GPT without apps in Instant Chat has no added cost, but creating and editing plugins through a conversation uses Work or Codex, and actions in Work consume credits. Heavier plugins have a running cost that the old GPT didn't make obvious, so check your usage after the first week.
Your time is the larger cost. These are estimates: 30 to 60 minutes for a GPT with instructions and files, including testing. Half a day for one with several connected apps. For a custom action, budget a few hours to a few days of developer time, depending on the API and its login method. This site's guide to AI automation agency pricing can help you judge a quote (builtwithagents.ai/blog/how-much-do-ai-automation-agencies-charge).
What breaks, and when not to bother
Start with whether the GPT deserves saving. If nobody has opened it in three months, let it retire. If it is a short prompt you use once a week, a saved skill or even a text snippet will do.
The most likely breakage is quiet: the plugin answers fine but stops following one instruction, or uses the wrong reference file. That's why you saved test prompts. The next one is access. Having permission to use an old GPT doesn't give anyone access to the new plugin, and included apps still need their own sign-in and approvals. Automatic selection is not guaranteed either, as OpenAI says the plugin won't necessarily run on every request. If a step must always happen, teach people to call it with @.
Agencies should watch one more trap. If you built GPTs inside client accounts, only the creator, or a workspace admin on Enterprise, can migrate them, so sort out who holds that access now. And public GPTs created in affected Enterprise workspaces follow that workspace's timeline even for outside users.
Frequently Asked Questions
When do custom GPTs retire?▾
OpenAI's help center lists December 11, 2026 as the scheduled retirement date. FedRAMP workspaces and Enterprise workspaces with a qualifying approved deferral have until February 11, 2027. Follow the notice for your own account, since OpenAI says the timeline communicated to you applies.
What happens to my GPT after December 11?▾
It and its page become inaccessible. Your old conversations with it remain accessible, and you don't need to do anything to preserve them. If you migrated it, you keep working in the plugin.
Do custom GPT actions transfer to plugins?▾
No. Custom actions don't transfer through migration. You need an existing app that covers the same job, or a rebuilt connection, which may require a custom MCP server and technical help. Instructions, knowledge files and connected apps do move.
What is the difference between a custom GPT and a plugin?▾
A GPT was a saved chat assistant with instructions, files and optional actions. A plugin bundles skills, connected apps and sometimes interactive views, and it works across ChatGPT Chat, Work and Codex where available. ChatGPT can also choose an installed skill automatically when your request matches its description.
Can I still create a new custom GPT?▾
Don't start one. OpenAI says new GPT creation ended on October 26, 2026 for affected Enterprise workspaces, except FedRAMP, and it points people to Plugin Creator for new reusable workflows. Even where creation still works, the GPT would retire in December.
What should I do if I only use someone else's GPT?▾
You can't migrate it, so check the notice on the GPT and ask its creator about a replacement. Access to the old GPT doesn't guarantee access to the new plugin. If the creator goes quiet, copy down how you use it and build your own skill before December 11.