Custom GPTs Going Away: Project, Skill or Plugin for Each One?

Coding Liquids tutorial cover featuring Sagnik Bhattacharya for Custom GPTs Going Away: Project, Skill or Plugin for Each One?
Coding Liquids tutorial cover featuring Sagnik Bhattacharya for Custom GPTs Going Away: Project, Skill or Plugin for Each One?

Custom GPTs are scheduled to retire on 11 December 2026, so decide each one separately: a GPT used for ongoing work with its own files becomes a ChatGPT project, a GPT that applies one repeatable format becomes a skill, and a GPT that connects to other services or serves a whole team becomes a plugin.

The Complete Excel Course with AI Integration

From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.

By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner

Buy now

Two facts shape every decision. OpenAI's migration turns a GPT's instructions into a skill and its knowledge files into reference files inside a new plugin, but custom actions, the chosen model, unpublished drafts and sharing settings do not come across. And chats you had with a GPT cannot be moved into a project, though they stay in your chat history.

When do custom GPTs stop working, and what happens on the day?

Custom GPTs are scheduled to retire on 11 December 2026 on every ChatGPT plan; on that date the GPTs and their GPT pages become inaccessible. FedRAMP workspaces and Enterprise workspaces with an approved deferral get until 28 February 2027.

OpenAI's custom GPT retirement and migration FAQ is the page to plan from, and it settles several questions owners ask:

  • Your GPTs work until the date. Existing GPTs stay usable until retirement, subject to their current access and workspace permissions.
  • Old conversations survive. Chats you had with a custom GPT stay accessible after retirement with no action needed.
  • Migration freezes the original. After you migrate a GPT, the original stays usable until retirement but becomes read-only, and its creator cannot delete it.
  • Timing depends on where the GPT was made. If you use several accounts or workspaces, the timeline of the workspace where the GPT was created applies, even if you use it from elsewhere.
  • Migration may arrive late. OpenAI says migration access may become available at different times, so do not leave everything until the week before.

One source of confusion: OpenAI's pricing page still lists custom GPTs among the Plus features. The help centre's retirement FAQ, and the notice in your own account, are what to plan around. On 22 November 2026, the standard date was 19 days away.

Project, skill or plugin: which replacement fits which kind of GPT?

A custom GPT's replacement follows how people use it: ongoing work with files suits a project, a repeatable task a skill, and connected services or a wide audience a plugin. What the GPT was called matters far less.

Use this table to sort each GPT. Read across the row that matches how people actually use it today.

If the GPT is mainly…Best replacementWhyWatch out for
Instructions plus files for one ongoing piece of work (a term plan, a client account, a research file)ChatGPT projectProjects keep chats, files and instructions together and can draw on earlier chats in the projectFile limits by plan; old GPT chats cannot be moved in
A house style or format used across many unrelated chats (emails, quotes, summaries)SkillChatGPT can apply an installed skill automatically when its description matches the requestStandalone skills are listed for Business, Enterprise, Healthcare and Edu users
Instructions that call connected apps, or a bundle of several workflows for a teamPluginA plugin packages skills and apps together and can be shared across a workspaceCustom actions need rebuilding; apps still need their own sign-in
A GPT with custom actions that reach your own systemsPlugin with a rebuilt app, or retire itActions do not transfer; rebuilding may need a custom MCP serverTechnical setup and a security review
A GPT your customers or students open by public linkSomething outside ChatGPT, or retire itMigration does not carry public sharing; public plugins need a separate submissionCustomers lose the link on the retirement date
Someone else's GPT that you only useNothing to migrateOnly the creator (or, in Enterprise, a workspace admin) can migrate itAccess to the GPT is not access to its replacement
A GPT nobody has opened in monthsLet it retireNo migration effort for no usersKeep a copy of its instructions text first

Two GPTs that look alike can land in different rows. A "client brief" GPT that one account manager uses for one big client is a project. A "client brief" GPT that six people use to turn any meeting note into the same brief format is a skill. Ask who uses it, how often, and whether the work is one thread or many.

What moves across when you migrate a GPT to a plugin?

OpenAI's built-in migration moves a GPT's published instructions, knowledge files and connected apps into a new private plugin. Custom actions, the selected model, drafts, public sharing and the GPT's existing users do not move.

Part of the GPTAfter migrationWhat you do about it
InstructionsBecome a skill within the new pluginReview the wording; test that ChatGPT picks the skill
Knowledge filesCopied into the plugin's reference filesCheck the plugin uses the right file for each answer
Connected appsAdded to the plugin as appsEach user still connects and authorises them
Custom actionsNot transferredFind an app that does the job, rebuild, or drop the feature
Selected modelDoes not carry overRetest anything that depended on a particular model
Drafts and unpublished editsNot transferred; migration uses the latest published versionPublish essential edits first (publishing need not mean sharing publicly)
Who could use itThe plugin starts private; existing users get no access automaticallyShare the plugin, then ask users to install and test it
Public sharingNot carried overA public personal plugin needs a separate submission process
Past conversationsStay in chat historyNothing; they cannot be moved into a project

One detail matters for GPTs that crunched numbers. OpenAI's FAQ says standard Chat, including Instant, does not currently run scripts bundled with installed plugins, and that after migration Chat may be able to read the reference files even when those files are not available to the tools that run code or generate files. A GPT that calculated fees from a spreadsheet in its knowledge files is the one to test hardest.

A music school's six GPTs, sorted in one afternoon

Here is an illustrative inventory for a music school with a ChatGPT Business workspace, four teaching staff with seats and one office manager. Between them they built six GPTs over two years. The owner blocked out an afternoon on 24 November to decide each one.

GPTUseFiles and toolsDecisionEstimated effort
Parent Email HelperOffice manager, about 30 emails a weekInstructions onlySkill (via migration)20 min migrate, 40 min test
Lesson PlannerAll four teachers, any student7 syllabus PDFsPlugin with skill and reference files, shared to the workspace30 min migrate, 60 min test
Term PlanningOwner only9 files: calendar, budget, room planProject25 min set-up
Fee CalculatorOffice manager, at enrolmentFee spreadsheet; worked out sibling discountsMove the calculation into the spreadsheet; retire the GPT90 min
Room Booking CheckerTeachers, a few times a weekCustom action calling the booking systemRetire; check the booking system directly10 min to tell staff
Theory Quiz BuddyStudents, by public linkInstructions and 3 worksheetsRetire; move quizzes to the school's own learning pages3 hours

The quick sum: roughly 8 hours 45 minutes of staff time across three weeks (20 + 40 + 30 + 60 + 25 + 90 + 10 + 180 = 455 minutes, plus about 70 minutes for a final round of user testing and a staff note). The two decisions that saved the most time were the two retirements. Rebuilding the booking action would have meant a developer and a security review for a check that takes 30 seconds in the booking system itself.

The Fee Calculator is the instructive one. It was really a spreadsheet wearing a chat interface: the GPT read the fee table and applied a 10% sibling discount. Given OpenAI's warning about code tools and reference files, the owner moved the discount rule into the spreadsheet as a formula and let the GPT go.

If building that fee formula, or checking the discount sums, is the slow part, try The Complete Excel Course with AI Integration on Coding Liquids.

How do you migrate a GPT to a plugin, step by step?

Migrating a GPT to a plugin starts from My GPTs: publish your final edits, select Migrate to plugin, follow the instructions, then review, test and share the private plugin before telling anyone to switch.

  1. Publish the final version. Open the GPT in the editor and publish any essential edits. Migration copies the latest published version only; publishing does not require sharing the GPT publicly.
  2. Save some test prompts. Copy three to five prompts your team uses often, plus one awkward case, and save the GPT's current answers to them. OpenAI suggests comparing familiar prompts and at least one harder case.
  3. Start the migration. Go to My GPTs and select Migrate to plugin for the GPT. If you do not see the option, migration may not be available to your account yet, plugin access may be off, or you may be in a different account or workspace from the one where you created the GPT.
  4. Review what was created. Read the migrated skill's instructions and check the reference files are all there.
  5. Test it. Run your saved prompts by mentioning the plugin with @, or by selecting + and entering its name. Compare against the saved answers.
  6. Fix and retest. Change the plugin, not the GPT, which is now read-only. To update a plugin, mention @plugin-creator with the plugin and describe the change.
  7. Share it. Go to Plugins, select the plugin, open the more options menu (•••), choose Share plugin, pick who has access and select Save. In a managed workspace, sharing can prefill the original GPT's audience; review it before saving.
  8. Switch users over. Ask one user to install and test it, then send everyone a short note with the date the old GPT stops.

Access options when sharing include Only those invited, Anyone in this workspace with the link, and Visible in workspace directory. Publishing to your workspace directory does not publish the plugin publicly. Workspace admins decide whether members can share or publish plugins at all, under Workspace settings, Permissions & roles, so check those settings before migration day if you are not the admin.

Where a migrated GPT went wrong: the unpublished fee change

An illustrative slip that is easy to make. The office manager updated the Parent Email Helper's instructions in October with the new term dates and the revised late-payment wording, but closed the editor without publishing. Migration copied the published version, so the new skill drafted emails with last term's dates. Because the original GPT was now read-only, the fix had to go into the skill. The lesson: publish, check the published GPT answers correctly, and only then migrate.

How do you rebuild a GPT as a ChatGPT project instead?

Rebuilding a GPT as a ChatGPT project takes four moves: select New project in the sidebar, paste the GPT's instructions into Project settings, upload its reference files, and start fresh chats inside the project. Custom actions and sharing settings do not come across.

  1. Copy the GPT's instructions text from the editor into a document, so you have it after retirement.
  2. Select New project in the sidebar, then name it.
  3. Open the more options menu (•••), select Project settings and paste the instructions. Project instructions apply only inside that project and override your general custom instructions.
  4. Upload the reference files. Only 10 files can be uploaded at once.
  5. Choose the memory setting. Under Project settings, Memory offers Default memory or Project-only memory; shared projects are always project-only.
  6. Start a new chat in the project and run your saved test prompts.

File and sharing limits differ by plan, which decides whether a GPT with many files fits in one project:

PlanFiles per projectCollaborators on a shared project
Free55
Go and Plus2510
Pro40100
Business, Enterprise and Edu40Up to 100 in a workspace project

Projects are included with every qualifying plan at no extra cost. Project sharing is available on Free, Go, Plus, Pro, Business, Enterprise and Edu. For setting up projects so each client's material stays apart, see using ChatGPT Projects to keep client work separate.

A before-and-after from the music school's Term Planning GPT. Before: the owner opened the GPT, re-explained which term she was planning, and pasted the latest room plan each time because the GPT's knowledge file was months out of date. After: a project called "Spring term 2027" with the instructions in Project settings, the current room plan and budget uploaded, and every planning chat inside it. A question such as "Which Tuesday slots are still free in Room 2?" now draws on the room plan and the earlier chat where she moved the piano exams, without re-pasting anything.

When is a standalone skill better than a full plugin?

A standalone skill beats a plugin when the GPT was only ever instructions and examples with no apps: a writing style, a document format or a checklist. Skills are listed for eligible Business, Enterprise, Healthcare and Edu users.

The difference in practice: a plugin is the container, and a skill is one set of reusable instructions inside it or on its own. OpenAI describes skills as reusable, shareable workflows that can include instructions, examples and code, and says ChatGPT can use an installed skill automatically when it helps. To find them, select Plugins in the sidebar, then the Skills tab in the Plugin Directory. Create a new one with Create, then Create with chat, Create with editor or Upload from your computer.

Because ChatGPT chooses a skill when its description matches the request, the description does more work than a GPT's name ever did. Here is the Parent Email Helper rewritten, illustrative.

Before, as GPT instructions: "You are Parent Pal, a friendly assistant for our music school. Help write emails to parents. Be warm but professional."

After, as a skill:

Description: Use when drafting or replying to emails to parents or guardians
about lessons, fees, exams, cancellations or practice at the music school.

Instructions:
- Open with "Dear [first name]," and sign off "Kind regards, [your name], Office".
- Keep emails under 150 words unless the parent asked several questions.
- Fees: quote only figures from the current fee sheet; never estimate.
- Cancellations: give the 48-hour notice rule and the make-up lesson option.
- Exams: never promise a grade or a pass.
- End with one clear next step for the parent.

Example request: "Reply to a parent asking to move Thursday's lesson."

The rewrite names when to use the skill, sets hard rules the old GPT left to chance, and includes an example request. When the office manager typed "draft a reply to a parent who missed two lessons" in an ordinary chat, the skill was picked up without being mentioned. Mention it with @ when you want to be certain.

On personal plans the options may differ: OpenAI says migration and plugin access vary by account or workspace and points users to the in-product notice. If your notice offers migration, take it; if not, a project holding the instructions is the safe fallback.

What happens to GPTs that use custom actions?

Custom actions in a GPT do not transfer through migration, so any feature that depends on one stops working in the plugin until you rebuild it with an app or a custom MCP server.

A custom action is the part of a GPT that called another system: checking a booking, creating a CRM record, looking up stock. OpenAI's advice is to check first whether an available app (formerly connectors) already supports the tasks you need. Rebuilding the connection may require a custom MCP server, which is a small piece of software that exposes your system to ChatGPT, and technical setup. Test any rebuilt integration before switching, and do not assume it does everything the original action did.

Questions to answer before paying anyone to rebuild an action:

  1. How many times a week does anyone use the action? The music school's booking checker ran a few times a week; that does not justify a rebuild.
  2. Does the other system already offer an app inside ChatGPT? If so, a plugin can include it.
  3. Does the system have its own AI assistant or search that does the same job?
  4. Who will maintain the rebuilt connection when the other system changes its API?
  5. What data will the connection be able to read and change, and who signs that off?

If you want help fitting the GPT replacements into how your business actually runs, I offer one-to-one AI consultations for small businesses.

What about a GPT your customers reach through a public link?

A public GPT that customers or students use has no automatic replacement: migration does not carry public sharing settings, and making a personal plugin public needs a separate plugin submission process. Plan for the link to stop working on the retirement date.

For a customer-facing GPT, ask what job it did. If it answered questions about your services, a chatbot on your own website is usually a better home, and website chat options as custom GPTs retire compares the routes. If it was a learning tool, like the music school's quiz GPT, the content can move to your own pages or a learning tool you control. Either way, tell users before the date. A suggested notice, filled in for the music school:

Subject: Theory Quiz Buddy is closing on 11 December

Dear [first name],

The Theory Quiz Buddy chat that many students use for grade practice will stop
working on 11 December 2026, when OpenAI retires custom GPTs.

From 1 December the same quizzes, with answers, will be on the Practice page of
our website. Nothing is needed from you; your child's past chats are not affected.

Kind regards,
[your name], Office

Also check whether any public GPT was built in an Enterprise workspace: OpenAI notes that public GPTs created in affected Enterprise workspaces follow the Enterprise transition even when people open them from personal accounts.

How do you test a replacement before 11 December?

Testing a GPT's replacement means running the same saved prompts through the old GPT and the new version, then checking skill choice, reference files, output format and tools.

Those four checks come from OpenAI's own FAQ. Turn them into a short test sheet per GPT. Here is the music school's, filled in for the Lesson Planner plugin (illustrative):

Test promptOld GPTNew pluginPass?Fix
30-minute Grade 3 piano lesson, scales focusWarm-up, scales, piece, homeworkSame structureYesNone
Grade 5 theory: requirements for the next examQuoted the syllabus PDFGave general advice, no syllabus referenceNoAdded "always check the syllabus files" to the skill
Lesson plan as a table for the parent portalFour-column tableBullet listNoAdded the table format as an example
Awkward case: adult beginner with arthritisShorter sessions, no medical adviceSameYesNone

A sample of the kind of difference to look for, from the second test (illustrative output from the new plugin before the fix): "For Grade 5 theory, students typically need to cover intervals, chords and transposition." That reads well but cites nothing. The old GPT had said which syllabus section applied. After adding one line to the skill telling it to check the syllabus reference files and name the section, the plugin's answer pointed to the right file again.

Test connected apps separately: OpenAI notes that apps inside a plugin still need their own access, account authorisation and action approvals, and installing a plugin grants no extra permissions in the connected service.

A three-week plan that finishes before the deadline

Working back from 11 December gives a calm schedule if you start in the last week of November. Leave the final week as a buffer, because migration access can arrive at different times for different accounts.

  • Week of 23 November: list every GPT you built or rely on, with who uses it, its files and whether it has actions or a public link. Make each decision using the table above. Copy every GPT's instructions into a document. Publish any unfinished edits.
  • Week of 30 November: set up projects, run migrations, rewrite skill descriptions, and run each test sheet. Send notices for public GPTs you are retiring.
  • Week of 7 December: share replacements, have each user install and test, fix the last failures and tell staff the old GPTs stop on 11 December.

Keep the inventory afterwards. It doubles as a record of which AI tools your business relies on, what files they hold and who maintains them, which is the list you need anyway when a staff member leaves or a supplier changes its terms. For what happens to files you uploaded to GPTs over the years, whether custom GPTs are private and what happens to uploaded files covers the privacy side.

Questions about the custom GPT retirement

Will I lose my old conversations with a custom GPT?

No. OpenAI's retirement FAQ says existing conversations with custom GPTs remain accessible after retirement and you do not need to do anything to keep them. What you lose is the GPT itself: on the retirement date, custom GPTs and their GPT pages become inaccessible, so you cannot start new chats with it. Those old chats also cannot be moved into a project.

Can I still edit a GPT after migrating it to a plugin?

No. Once you migrate, the original GPT stays usable until retirement but becomes read-only, and its creator cannot delete it. Make every essential edit and publish it before you start migration, because the migration copies the latest published version, not drafts. After that, all changes go into the plugin, which you update through Plugin Creator.

Do I have to migrate a GPT someone else built?

No. OpenAI says you do not need to migrate a GPT you only use, and permission to use it does not let you migrate it. Check the notice on the GPT and any guidance from its creator about a replacement. Access to the original GPT does not automatically give you access to the replacement plugin, so confirm you can open the replacement before the retirement date.

Is there a later retirement date for some workspaces?

Yes, for a small group. FedRAMP workspaces and affected Enterprise workspaces with a qualifying, approved deferral retire on 28 February 2027 instead of 11 December 2026. OpenAI says the later date does not apply to every workspace and that a deferral does not keep the ability to create new GPTs. Follow the notice in your own account or workspace.

Why does OpenAI's pricing page still list custom GPTs on Plus?

The pricing page still lists custom GPTs under Plus and Pro, but OpenAI's help centre is the source for the retirement: custom GPTs are scheduled to retire on 11 December 2026 across all ChatGPT plans. Plan around the retirement FAQ and the notice in your account, not the plan feature list, which can lag behind product changes.

Further reads

Sources: OpenAI help centre articles Custom GPT retirement and migration FAQ, Plugins in ChatGPT, Skills in ChatGPT and Projects in ChatGPT; ChatGPT pricing page. All read 22 November 2026.

Want a second pair of eyes on your setup?

Bring the process you're trying to fix. On a 1:1 call we'll work out which AI tools should handle it, which parts stay manual, and how to roll it out without disrupting the team.

Book a 1:1 call with me