Custom GPTs from 11 December 2026. Plugin or project: where to move them, and what you lose either way.
On 11 December 2026 OpenAI retires custom GPTs and offers to migrate them to plugins; for everyday work it points to Projects. Here is how to choose between the two routes, what does not carry over, and a checklist in Word and PDF.
OpenAI has set the retirement of custom GPTs for 11 December 2026, on every plan, and is helping creators migrate them to plugins. From that day GPTs and their pages become inaccessible; past chats stay readable.
There are two routes. The plugin is the official migration, when your account offers it: the instructions become a skill. The project is the alternative OpenAI recommends for ongoing work; it runs on every plan and is built by hand.
Neither route carries over the selected model, custom actions or sharing settings. The chats you had with the GPT stay in your history.
The checklist for choosing and moving can be downloaded below, in Word and PDF, with no email required.
Written by: Mattia Esposito, AI and marketing consultant, founder of Itria AI. From Puglia, based in Turin since 2023, he works remotely in Italy and abroad. LinkedIn profile.
The custom GPT checklist: plugin or project, to download
A six-part checklist for every GPT you use, before 11 December 2026: inventory, what to copy, the choice between plugin and project, the steps for each route, the test, what does not carry over. Fill it in on a computer, because the GPT editor lives on the web. OpenAI advises allowing extra time for GPTs with custom actions, the ones that connect them to other software.

| File | Contents | Link |
|---|---|---|
| Wordto fill in, in Italian | The six parts, with one row per GPT and boxes to tick. | |
| PDFto print, in Italian | The same checklist, to keep beside you while you copy. |
| Part | What you do | Where it comes from |
|---|---|---|
| 1. Inventory | Every GPT you use: who created it, in which account or workspace, who uses it. | OpenAI's retirement FAQ |
| 2. What to copy | Name, full instructions, knowledge files, conversation starters, model, capabilities, actions, sharing. | The fields of a GPT (OpenAI) |
| 3. The choice | Plugin, project, or nothing if the GPT belongs to someone else, with the questions that decide it. | OpenAI's retirement FAQ |
| 4. The steps | For the plugin: review the skill and files, share it again. For the project: instructions, files, memory, access. | Help Center pages on plugins and Projects |
| 5. The test | Three requests you used to give the GPT and one hard one, with the answer before and after. | The test OpenAI asks for |
| 6. What does not carry over | Selected model, custom actions, sharing, public link; old chats stay in your history. | OpenAI's retirement FAQ |
On 5 October I wrote about how Itria began. Seven months earlier, for a marketing exam, I had asked Gemini to link every topic on the syllabus to some aspect of the business I had in mind. Then I spent three months building the first version before putting it in front of anyone.
In the post I listed it among the things not to repeat, with a rule: “Once you have solved the problem, freeze the build and move on to validating” (the full post, in Italian).
A GPT's retirement is a good moment to apply it. Move what you actually use: the instructions that gave good answers and the files the GPT drew on. Features you added and never used can stay where they are, and the test in part 5, with the requests you made every week, shows whether anything is missing.
The file. The guide to brand tone of voice in ChatGPT and Claude offers an eight-part voice document: who is speaking and to whom, the traits of the voice, tone by channel, the rules, the constructions to cut, the approved texts. It is typical material for a writing GPT: rules plus reference documents.
Where it goes now. In a plugin, the rules become a skill and the document a reference file. In a project, the short rules go in the instructions and the full document as a file; the instructions apply only there, so the company's voice stays out of your other chats.
Which route. For a team writing together, the shared project: OpenAI names this very use, project instructions to keep tone and style consistent across contributors.
Custom GPTs: what they are, and what happens on 11 December 2026
A custom GPT is a version of ChatGPT with instructions, knowledge files and tools chosen by its creator, used as it is by anyone who opens it. It has a name, a description, conversation starters, up to 20 files of 512 MB, a recommended model and, optionally, apps or actions connecting it to external services.
OpenAI announced the retirement of custom GPTs on 11 September 2026. Its FAQ, updated on 12 November, puts it like this: “Custom GPTs are scheduled to retire on Dec 11, 2026.” It applies to every plan, and on the retirement date GPTs and their pages become inaccessible.
| Date | What happens | For |
|---|---|---|
| 11 September 2026 | Notice to administrators and announcement of the retirement. | Enterprise workspaces; public announcement |
| Already today | No new GPTs and no publishing; existing ones can still be used and edited. | Free, Go, Plus, Pro |
| 26 October 2026 | New GPT creation closed; existing GPTs that have not been migrated can still be edited. | Affected Enterprise workspaces, except FedRAMP |
| 11 December 2026 | Retirement: GPTs and GPT pages become inaccessible. | All plans |
| 11 February 2027 | Later retirement. | FedRAMP workspaces and Enterprise workspaces with an approved deferral |
Three things remain. Chats you have already had with a GPT stay readable after the retirement, with nothing to do on your part. A GPT that has been migrated stays usable until the date, read-only, and its creator can no longer delete it. For Business workspaces OpenAI gives no date of their own: the standard date and the workspace notice apply.
The date can differ for a single account: the FAQ always points to the notice shown in the product, and it was updated twice in three days. The English-language tech press has followed the story, for example Virtualization Review on 28 September, which also reports pushback from some GPT builders: in their view a plugin does not reproduce the dedicated, predictable environment of a GPT.
Plugin or project: how to choose
OpenAI's FAQ separates the two cases: the plugin is the migration for the GPT's creator, the project is the alternative “for ongoing work, like writing, research, or planning”. The deciding question is what you used the GPT for: a reusable tool, perhaps connected to other services, or work that carries on and grows with chats and files.
| What | Plugin (migration) | Project |
|---|---|---|
| Who can do it | The GPT's creator, when the account shows the option; in an Enterprise workspace, an admin too. | Anyone, on every plan |
| Instructions | They become a skill, automatically. | Pasted into the project settings |
| Knowledge files | Copied as reference files. | Uploaded again, from 5 to 40 depending on the plan |
| Connected apps | Carried over as the plugin's apps. | Used from the tools menu |
| Custom actions | Not carried over. | Not carried over |
| Selected model | Not carried over. | Chosen in the chat |
| Sharing | The plugin starts private and has to be shared again. | Shared from scratch, with chat or edit access |
| Right for you if | The tool gets reused across different chats, outside a project too. | The work carries on and builds up over time |
One limit matters on personal plans. OpenAI's plugins page documents personal skills as available on Business, Enterprise, Healthcare and Edu, and says nothing of the kind for Plus; the FAQ adds that migration and plugin access can differ from one account to another. On Plus, if the option does not appear, the project is the safe route.
Migrating a GPT to a plugin: what carries over and what does not
The official migration turns a GPT into a plugin: the instructions become a skill, the knowledge files become reference files, connected apps come across as apps. You start it from My GPTs with Migrate to plugin, when your account offers it, and it uses the GPT's latest published version. Drafts do not transfer; publishing does not mean making it public.
The selected model, the sharing settings and custom actions are left out. For actions, OpenAI writes that rebuilding them may require a custom MCP server and technical setup. MCP, the Model Context Protocol, is an open standard for connecting AI applications to external systems: its documentation compares it to a USB-C port.
The plugin starts private: people who used the GPT do not get it automatically, and it has to be shared again. The original GPT stays usable until retirement, read-only, so make your last edits before migrating. ChatGPT can also use an installed skill on its own, when the skill's description matches the request.
If you use someone else's GPT, you have nothing to migrate and cannot do it anyway: check the notice on the GPT and what its creator says. Public GPTs created in an Enterprise workspace follow the Enterprise timeline even when you open them from a personal account.
ChatGPT Projects: what they are and what they are for
A ChatGPT project is a workspace that keeps chats, files and instructions together, and ChatGPT uses that context in every new chat started inside it. OpenAI's Help Center page, updated on 11 November 2026, recommends them for ongoing work: writing, research, planning.
Projects are unlimited and included in your plan: asked whether they cost extra, OpenAI answers “No”. You can start on one device and carry on with another. Inside a project you keep the usual tools, from web search to images, and a good response, such as a summary or a draft, can be saved to the project sources and reused.
OpenAI sets them apart from GPTs in a table: the GPT is “single player” and static, the shared project is multi-player and changes with the group's contributions. You can still call a GPT in a chat already open in the project; if the first message uses a GPT, the chat starts outside, and chats with a GPT cannot be moved in.
For a small business the typical uses are the recurring ones: proposals for a client, the weekly report, product sheets. How to do each of these well is in the guide to using ChatGPT for work. When the work means acting on other software, a project is no longer enough, and the worksheet for designing an AI agent shows where one ends and the other begins.
How to create a project in ChatGPT, and how it works
You create a project from the ChatGPT sidebar with New project: a name, an icon and a colour, then instructions and files. Instructions go in Project settings, from the three-dot menu; files are uploaded from the sources area. Chats you already have can be dragged in, except those started with a GPT.
If the project is taking a GPT's place, this order works. Open the GPT in the editor, on the web, and copy the text of its instructions into a document. Find the original knowledge files on your computer. Create the project, paste the instructions, upload the files. Then open a new chat in the project and repeat the requests you used to give the GPT, one by one.
OpenAI gives advice on writing instructions that applies to projects as well: positive, concrete instructions instead of long lists of prohibitions, rules and tone in the instructions, reference material in the files. The GPT's conversation starters have no field in a project, so keep them in a text file among the sources.
Project instructions and files: the limits per plan
Project instructions apply only inside the project: “Project instructions apply only within that project and override your global custom instructions”, OpenAI writes. Files have a ceiling that changes with the plan, and you can upload at most ten at a time.
| Plan | Files per project and collaborators | Is the content used to train the models? |
|---|---|---|
| Free | 5 files, 5 collaborators. | Yes, if “Improve the model for everyone” is on |
| Go and Plus | 25 files, 10 collaborators. | Yes, if the setting is on |
| Pro | 40 files, 100 collaborators. | Yes, if the setting is on |
| Business | 40 files, up to 100 collaborators; invitations to groups or the whole workspace. | No, by default |
| Enterprise and Edu | 40 files, up to 100 collaborators. | No, by default |
If an upload stops, there are two documented causes. You have reached the file limit: OpenAI suggests removing old files, combining their contents or splitting the work across several projects. Or you are uploading a file with the same name as one already there, and ChatGPT asks whether to Upload anyway or Skip it. Drive and Slack links have a separate limit.
Before uploading, one question: what is in those files? A project on a personal plan with the training setting on can help improve the models. Confidential price lists, contracts and client data belong on a business plan, and companies still using personal accounts for work are the case described in the glossary entry on shadow AI.
Sharing a project on Business plans, and where the data goes
On ChatGPT Business a project can be shared with people, groups or the whole workspace, up to 100 collaborators, with chat or edit access. Anyone joining through a workspace link starts with chat access. Members with edit access can change instructions and files and invite others, but cannot remove anyone.
The moment you share it, a Business project switches to project-only memory: it cannot see members' memories or personal instructions. Every member can download the project's files, so a file uploaded there is a file handed to each of them. A deleted project is removed from OpenAI's systems within 30 days.
On data, OpenAI writes that by default it does not train its models on Business content, and it will sign a Data Processing Addendum, the contract that Article 28 of the GDPR requires when a supplier processes personal data on your behalf.
Messages in a project inherit the workspace's retention and data region. Data residency in Europe, on the OpenAI page read on 13 November, is listed for API customers and new Enterprise and Edu customers; Business is not on the list. The full picture, CLOUD Act included, is in the guide to ChatGPT, the GDPR and the CLOUD Act.
Beyond ChatGPT: Claude projects and Gemini Gems
Google is also winding down its custom assistants: from November 2026 Gems on personal Google accounts move to skills, and from 2027 Gems on work and school accounts follow. The difference from OpenAI lies in how: Google writes that it will recreate Gems as skills automatically, and you can do it earlier by hand if you prefer.
Anthropic offers projects in Claude: spaces with their own chats, documents and instructions, available on the free plan too, up to five projects. A new version is in beta, starting with Claude Code, and existing projects of some Pro, Max and Ultra users are being moved across with nothing lost, instructions and history included.
The voice document and the instructions you copied from the GPT work there too: they are just text. How to use Claude in a company, plan by plan, is in the guide to using Claude for work; the comparison of the three assistants by task is in ChatGPT, Claude or Gemini for work.
The other guides on automation and AI at work, together with the marketing ones, are gathered in the index of guides for small businesses. The systems Itria builds on top of these tools, connected to a company's software and with a person approving, are on the custom systems page.
Questions and answers
What happens to custom GPTs on 11 December 2026?
OpenAI retires them: from 11 December 2026 custom GPTs and their GPT pages become inaccessible, on every plan. Chats you have already had with a GPT stay readable in your history.
Only FedRAMP workspaces and Enterprise workspaces with an approved deferral get 11 February 2027. The notice in your own account is what counts.
Where should a custom GPT go: a plugin or a project?
To a plugin, if you created it and your account shows the migration option: the official route, where the instructions become a skill. To a project, if you used it for ongoing work such as writing or research, or if the option is not there: Projects work on every plan, the free one included.
Do custom GPTs turn into projects automatically?
No. The migration OpenAI documents leads to a plugin, and the GPT’s creator starts it when the account offers it. A project is built by hand: you copy over the instructions and files. Either way, custom actions do not carry over and have to be rebuilt.
Can you still create a custom GPT?
Not on a personal account: Free, Go, Plus and Pro can no longer create or publish new GPTs, though they can still use and edit the ones they have. In affected Enterprise workspaces creation closed on 26 October 2026; on Business, the workspace notice decides. For new work OpenAI points to plugins and projects.
Do ChatGPT Projects work on the free plan?
Yes. Projects are available on Free too, with a limit of 5 files per project and 5 collaborators if you share it. Go and Plus allow 25 files; Pro, Business, Enterprise and Edu allow 40. Files upload ten at a time.
Notes on sources
- OpenAI, Help Center, Custom GPT retirement and migration FAQ, updated on 12 November 2026: retirement dates, migration to plugins, the project alternative, the Enterprise timeline; both quoted sentences come from here. Reread on 13 November 2026. The page changes often: the notice in your account is what counts.
- OpenAI, Help Center, Creating and editing GPTs, updated around 3 November 2026: the end of creation on personal accounts, the fields of a GPT, the limit of 20 files of 512 MB, advice on instructions.
- OpenAI, Help Center, Projects in ChatGPT, updated on 11 November 2026: file and collaborator limits, instructions (the quoted sentence comes from here), memory, sharing, data, the table comparing shared projects and GPTs. Read on 13 November 2026.
- OpenAI, Help Center, Plugins in ChatGPT, updated around 28 October 2026: what a plugin contains and the availability of personal skills.
- OpenAI, Enterprise privacy (training and the Data Processing Addendum) and Help Center, Data residency and inference residency for ChatGPT, updated on 10 November 2026. Read on 13 November 2026.
- Anthropic, Claude Help Center, What are projects?; Google, Gemini Apps Help, About the transition from Gems to skills. Read on 13 November 2026.
- Model Context Protocol, introduction, for the definition of MCP. Virtualization Review, OpenAI to Retire Custom GPTs, Replace Them with Plugins, 28 September 2026, for the date of the announcement and the GPT builders' criticism.
- Regulation (EU) 2016/679, text on EUR-Lex, Article 28. The post of 5 October 2026 is a public LinkedIn post by Mattia Esposito, linked in the text.
Choosing between plugin and project takes the checklist. Connecting that work to the company's software, with a person approving before anything goes out, is the next step. The system first, then the growth.
Mattia Esposito, who wrote this guide, builds tailored AI systems for small businesses with Itria: automations connected to the software the business already uses, and a Dashboard showing what has happened and what is waiting for a decision. Nothing goes to a customer without a person’s yes. On a test bench on 4 September 2026, with 30 documents including orders, delivery notes and invoices, the system extracted 234 fields out of 237 correctly, and the unreadable document deliberately placed in the batch was set aside without a single invented field.