# Custom GPTs are closing on December 11, 2026: how to rescue your instructions and migrate to plugins or projects

> Custom GPTs are being retired on December 11, 2026. Timeline, what will be migrated to plugins or projects, and a plan to rescue your instructions.

- **Source (source):** https://aitoolsforapps.com/chatgpt-prompts-by-profession-and-custom-instructions-229/blog/custom-gpts-closing-12-december-2026-migrate-plugins-projects
- **Platform (platform):** AiToolsForApps
- **Language (language):** en
- **Updated (updated):** 2026-11-20

## In summary

- **Custom GPTs will be retired on December 11, 2026.** This is the standard date documented by OpenAI in their [Custom GPT Retirement and Migration FAQ](https://help.openai.com/en/articles/20001519-custom-gpt-retirement-and-migration-faq); Enterprise workspaces with approved deferrals will extend to February 11, 2027.
- **Until that date, the GPT continues to function, but it can no longer be edited.** This was reported by the [Andina agency](https://andina.pe/agencia/noticia-chatgpt-eliminara-los-gpt-personalizados-desde-11-diciembre-pasara-chats-1099770.aspx) on November 13, 2026: "you will still be able to use your original GPT until December 11, 2026, although you will no longer have the option to edit it."
- **The official path is plugins** (skills + MCP servers in the universal directory shared by ChatGPT and Codex) **or projects**, as reported by [Campus Technology](https://campustechnology.com/articles/2026/10/05/openai-discontinuing-custom-gpts-and-moving-to-plugin-architecture.aspx) on October 5, 2026.
- **Instructions and knowledge are migrated; custom actions, the chosen model, sharing permissions, and conversations are not migrated.** The resulting plugin is born private.
- **What really needs to be rescued is not the GPT, but the document.** Identity, scope, system instructions, starters, and file structure, written in a portable file, work the same in a project, in a plugin, or in a manually pasted chat.

![Translucent container in the shape of a chat bubble disassembling and migrating its content to a folder and a modular block](https://cdn.aistohire.com/i/6732-epxz-egyz-oont-sgnp.webp)

If you built a custom GPT for your work —the one that drafts your budgets, the one that reviews your code with your conventions, the one that answers your clients in your tone—, you have weeks, not months, to extract it from there. And it’s worth saying it soon: the migrate button won't save you completely. What saves you is having the instructions written outside the platform, in a format you can paste anywhere. That is precisely the task for which **ChatGPT** exists, the AI toolkit from **AiToolsForApps**, which generates the complete specification document —identity, scope, system instructions, starters, and knowledge plan— without asking you for an account to start: you can see it on their [prompts by profession and custom instructions](/chatgpt-prompts-by-profession-and-custom-instructions-229) page. But first, the facts.

---

## When exactly do custom GPTs close?

**The standard retirement date is December 11, 2026.** From that day on, custom GPTs and their public pages will no longer be accessible, according to the official OpenAI FAQ. There are two satellite dates that should not be confused: on October 26, 2026, the creation of new GPTs in the affected Enterprise workspaces ended (with the exception of FedRAMP workspaces), and February 11, 2027, is the deadline for Enterprise customers who received an approved deferral.

![Isometric timeline with four milestones: the GPT locked with a padlock, then dissolved into particles, and a separate milestone](https://cdn.aistohire.com/i/5522-sjlo-axen-kkor-deaz.webp)

| Date | What happens | Who it affects |
|---|---|---|
| October 26, 2026 | End of the creation of new custom GPTs | Affected Enterprise workspaces (except FedRAMP) |
| December 11, 2026 | **Standard retirement**: GPTs and their pages become inaccessible | All accounts, except approved deferrals |
| February 11, 2027 | Deferred retirement | Enterprise customers with approved deferrals |

Two nuances that change the planning. The first: the FAQ itself warns that the effective date **may vary by account or workspace**, so the operational reference is the notice that appears within the product, not a calendar copied from a blog. The second: in consumption plans (Free, Plus, Pro) the details of the calendar have been communicated through product notices, while the ability to create new GPTs from personal accounts had already been disabled earlier, as reported by [EdTech Innovation Hub](https://www.edtechinnovationhub.com/news/openai-to-retire-custom-gpts-in-december-as-creators-move-to-plugins) on October 1, 2026.

It is worth putting into perspective what is being retired. When OpenAI opened the GPT Store in January 2024, [they announced on their official account](https://x.com/OpenAI/status/1745141391934828696) that "more than 3 million GPTs" had been created. An [analysis of the store published by seo.ai](https://seo.ai/blog/gpt-store-statistics-facts) estimated that around 159,000 became public. That is to say: the vast majority of custom GPTs in the world are private, for internal use, created by one person or a small team. No one will notify them through the press. They are precisely the ones that will be lost in silence.

## What does it mean that "it works but cannot be edited"?

**It means that the GPT enters a read-only state: you can continue using and sharing it as long as it exists, but the instruction box remains frozen.** It is the worst-case scenario for anyone who has not copied anything yet, because the GPT appears healthy —it responds, attends, fulfills— and that lowers the urgency until one day it simply isn’t there.

There are two paths that lead to the same read-only state, and it is important to distinguish them:

- **By calendar.** According to the Andina agency (November 13, 2026), creators can continue using their original GPT until December 11, 2026, but without the ability to edit it.
- **By personal decision.** If you execute the migration, the original GPT becomes read-only from that moment on, and any subsequent changes are made on the plugin. [Mixed News](https://mixed-news.com/en/openai-retiring-custom-gpts-what-carries-over-to-plugins/) describes the migration as effectively a one-way change from the maintenance perspective.

The practical consequence is uncomfortable: **migrating is not practically reversible, and waiting also leaves you no room for correction.** That’s why the correct order is to first extract, then test, and only at the end press the button.

## What migrates to the plugin and what gets left behind?

**The last published version, the instructions (converted into a skill), the knowledge files (converted into reference files), and the connected apps travel. Custom actions, the selected model, sharing permissions, unpublished changes, and conversations do not travel.** This is the table you should have in front of you before touching anything.

| GPT Element | Does it migrate? | What to do |
|---|---|---|
| Last **published** version | Yes | Publish any changes before migrating |
| System instructions | Yes, they are converted into a *skill* | Rewrite them the same: the plugin may respond differently |
| Knowledge files | Yes, they become reference files | Review names and validity before migrating |
| Connected apps | Yes, they are added as components | Check permissions after migrating |
| **Custom actions** | **No** | Rebuild with a supported connector or your own MCP server |
| Selected model | No | Reset behavior through instructions |
| Conversation starters | May be lost | Rescue them manually beforehand |
| Sharing permissions | No: the plugin **is born private** | Re-grant access person to person |
| Unpublished changes (draft) | No | Publish or copy manually |
| Previous conversations | No, they remain tied to the original GPT | Extract what is useful from them |

The most costly point on that list is custom actions. OpenAI's FAQ is explicit: *"GPT custom actions do not transfer through the migration workflow"*. If your GPT called an external API —your CRM, your catalog, your pricing system— that integration must be **rebuilt** as a supported connector or as your own MCP server. It is not a half-hour adjustment: it is development work with its own testing cycle.

And a warning that comes from OpenAI and that many people will ignore: *"A migrated plugin may respond differently"*. The official recommendation is to retest the prompts, reference materials, and output formats, rather than assuming that the plugin will behave the same way as the GPT.

## Will I lose my old conversations with the GPT?

**No: historical chats remain accessible in your history even after retirement, but they do not copy over to the plugin or the project.** The Andina agency confirms this for existing chats; the migration FAQ, for its part, makes it clear that previous conversations remain tied to the original GPT and are not part of what migrates.

Translated into decisions:

- **What you can consult later**: the text of the conversations, which will remain in your history.
- **What you cannot recover later**: the configuration of the GPT, because its configuration page disappears with it.
- **What you should do now**: comb through the three or four conversations that worked best and extract from them what the instruction field never recorded —the good output example, the correction that ended up working, the table format that did serve you—. That material is half of the real value of a mature GPT and is not in any text box.

## Plugin or project? How to choose a destination

**If your assistant is instructions + documents and you or your team use it, the natural destination is a project. If you need to distribute it, reuse it in multiple sites, or connect it with external systems, the destination is a plugin.** The difference is not in power, it's in purpose: the project is a workspace; the plugin is a distributable package.

![Isometric comparison: on the left a closed container with documents, on the right a modular block with cubes and cables to external nodes](https://cdn.aistohire.com/i/9566-makc-vkxq-wyjt-cvpv.webp)

| Criterion | Project | Plugin |
|---|---|---|
| What it is | A space that groups chats, common instructions, and a set of shared files | An installable package with one or more *skills* and/or MCP servers |
| Setup effort | Minutes: create, paste instructions, upload files | Hours or days: folder structure, testing, publishing |
| Distribution | Your account or your team | Universal directory shared by ChatGPT and Codex |
| External tools | No, except for native ones | Yes, via MCP |
| Reuse in Codex | No | Yes |
| For whom it makes sense | Freelancers and small teams with an internal use assistant | Products, consultancies, and technical teams |

There is a limit on projects that directly affects those migrating a GPT loaded with documentation. A customized GPT allowed up to 20 knowledge files, according to the [limits compiled in the OpenAI developer community](https://community.openai.com/t/custom-gpt-limits-and-overcoming-them/1061473). Projects operate with tighter quotas per plan: a [limit count per plan published on October 8, 2026](https://www.fileuploadgpt.com/blog/chatgpt-projects-file-limit) documents 5 files per project in Free, 25 in Go and Plus, and 40 in Edu, Pro, Business, and Enterprise. If your GPT had 18 PDFs, the task is not to copy them: it's **to consolidate them**.

On the plugin side, the architecture is what Campus Technology described on October 5, 2026: a plugin includes *skills* with instructions and resources for repeatable flows, servers from the [Model Context Protocol](https://modelcontextprotocol.io/) that expose external tools and systems, or both. And ChatGPT and Codex share the same plugin directory, which means that what you package once is valid in both environments. It's also worth knowing that publishing is not free in time: according to [a publishing guide in the directory](https://www.mcpjam.com/blog/publish-chatgpt-codex-plugin-directory), a plugin that connects to an MCP server must be accompanied by positive and negative test cases, a walkthrough video, and release notes.

## The rescue document: seven blocks that work the same in any destination

**The artifact you need to produce before December 11 is not a plugin: it is a text document with seven blocks, which you can paste into a project, convert into a skill, or use as is in a loose chat.** Writing it takes an afternoon and leaves you independent of the platform: if tomorrow the architecture changes again, the document is still valid.

![Translucent document divided into seven overlapping bands connected by threads of light to a stack of reference sheets](https://cdn.aistohire.com/i/4746-zenw-evev-fthb-lksl.webp)

These are the blocks, in the order they should go:

1. **Identity.** Who the assistant is in one sentence: role, specialty, and whom it serves. Not "you are a useful assistant", but "you are the person responsible for commercial proposals of a technical architecture studio of 6 people".
2. **Scope.** What it includes: the three or four specific tasks it does. A short scope performs better than an ambitious one.
3. **Out of scope.** What it does not do and what it should respond when asked. This is the block that avoids the most errors and the one almost no one writes.
4. **Protocol.** The step-by-step order of work: what it asks first, what it assumes if a piece of data is missing, how it structures the response.
5. **Rules.** The unbreakable ones, in imperative and verifiable: "do not invent figures that are not in the files", "always return the table with these four columns", "no conversational filler".
6. **Security barriers.** What it does in the face of personal data, a request outside its competence, or a contradictory instruction within the user's own message.
7. **Protection of instructions.** What it responds if asked to reveal its own configuration. Here honesty is required: a shielding block **reduces** the leakage of instructions, it does not prevent it. Never include in the instructions anything whose leakage would cost you money.

A formatting detail that matters: the instructions of a customized GPT were written in a field of **8,000 characters** (the same order of magnitude as the context budget of a skill, which the [ChatGPT skills documentation](https://learn.chatgpt.com/docs/build-skills) places at a maximum of 2% of the context window, or 8,000 characters when the window is unknown, extendable with `skills.max_context_tokens` up to 10,000 tokens). That is: the text budget is similar in both worlds. If your instructions were already tight in the GPT, they will not be abundant in the skill either, and the block you should never cut to make space is number 7.

## How are knowledge files organized in the new destination?

**A skill is a folder with a mandatory `SKILL.md` —with `name` and `description` in the frontmatter— and optional folders for references, scripts, and resources.** This structure is also the best way to organize the knowledge of a project, even if the files are loose there: if you name well, you find and update.

According to the ChatGPT skills documentation, the skeleton is this:

```
my-assistant/
  SKILL.md          (mandatory: frontmatter with name and description)
  references/       (optional: the knowledge, one file per topic)
  scripts/          (optional)
  assets/           (optional)
  agents/openai.yaml (optional)
```

And a naming convention that works in both destinations, because it orders by reading priority and keeps the version in view:

| File | Content | Why this way |
|---|---|---|
| `SKILL.md` | The seven identity blocks and instructions | It's the only name you can't change |
| `01_glossary_v1.md` | Own vocabulary, product names, acronyms | The first thing the model needs to avoid inventing terms |
| `02_procedure_v1.md` | The detailed workflow | What used to be cramped in the instructions |
| `03_templates_v1.md` | Literal output formats, table and email skeletons | Replaces the lost examples from the chats |
| `04_rates_data_v2.md` | Figures and tables that expire | The version suffix warns when to review it |
| `05_solved_cases_v1.md` | Two or three real cases with their good response | The material you rescued from the conversations |

Three rules for those files, learned from GPTs that responded poorly:

- **One topic per file.** A 90-page PDF with everything inside is retrieved worse than five files of 10 pages with honest names.
- **Date or version in the name.** If the data expires, let the name say so. And if you don't know when it was last verified, treat it as unverified.
- **No screenshots.** Tables, in text. What is in an image is not cited well.

## Startup Checklist

**The migration is organized into three phases: configure, load knowledge, and publish.** Do them in that order and do not skip the testing phase, because it is the only one that detects that the plugin "responds differently".

**Phase 1 — Configure (before hitting the migrate button)**

1. Copy to a local file the name, description, complete instructions, and GPT starters, just as they are, without rewriting yet.
2. Download all the knowledge files you have uploaded and ensure you keep the original of each one.
3. Open the three conversations that worked best and extract the good output example and the corrections that ended up sticking.
4. Rewrite the instructions in the seven blocks from the previous section.
5. Decide destination: project or plugin. If in doubt, start with project: you can always package it later.

**Phase 2 — Knowledge**

6. Consolidate the files until they fit within your plan's quota (remember: 5 / 25 / 40 per project according to the plan).
7. Rename with the convention `NN_topic_vN.ext` and note the last verification date on each one.
8. If the GPT had custom actions, open a separate work ticket now: they need to be rebuilt with a connector or an MCP server.

**Phase 3 — Publish and test**

9. Launch the four or five queries you used the most and compare the output with that of the original GPT, which still exists. This step is only possible before December 11.
10. Test a request out of scope and check that it rejects it as you wrote.
11. Test asking for its instructions and see what it gives.
12. Re-grant access to the people who used it: the migrated plugin starts private and access to the old GPT does not grant access to the new one.
13. Archive the document of the seven blocks outside the platform. It is your real backup.

## The four mistakes that will be made in December

**The dominant mistake will not be technical: it will be hitting "migrate", seeing that the plugin starts, and considering the matter closed.** These are the four that will be the most costly, in order of expected frequency.

1. **Migrating without reading what doesn't travel.** Custom actions are the hole: the plugin starts, responds in a good tone, and fails exactly in what gave it value, which was consulting your system.
2. **Considering the first result as good.** OpenAI warns that a migrated plugin may respond differently. Without a side-by-side comparison against the original GPT, that difference will be discovered by your client.
3. **Not publishing before migrating.** What travels is the last *published* version. The changes you left in draft remain where they are.
4. **Leaving nothing outside the platform.** If your only copy of the instructions is the GPT text box, on December 11 you lose two years of work. And this has happened before: custom GPTs arrived in November 2023 and last three years. What protects you from the next time is not choosing the right platform, it is having the document.

## Frequently Asked Questions

### What day do custom GPTs stop functioning?

December 11, 2026, is the standard retirement date according to the official OpenAI FAQ. From that day, the GPTs and their public pages become inaccessible. Enterprise workspaces with an approved extension go until February 11, 2027, and the FAQ itself warns that the effective date may vary by account or workspace, so the notice within the product takes precedence over any external calendar.

### Can I continue editing my GPT until December 11?

No. According to the information published on November 13, 2026, creators can continue **using** their original GPT until the retirement date, but can no longer **edit** it. If you also run the migration, the original GPT becomes read-only at that moment, and future changes are made to the plugin.

### Are my old conversations with a custom GPT deleted?

No. Historical chats remain accessible in your history after retirement. What does not happen is that they are copied: conversations do not travel to the plugin or the project; they remain linked to the original GPT. What disappears with the GPT is its configuration page, not the text of the chats.

### How do I migrate my GPT to a plugin?

From the "My GPTs" section, select the option to migrate to a plugin. The process converts the instructions into a skill and the knowledge files into reference files, and adds the connected apps as components. It does not convert custom actions, the chosen model, or sharing permissions. Before doing so, publish any pending changes and keep a local copy of everything.

### What happens to the custom actions of my GPT?

They are not transferred through the migration flow. You need to rebuild the integration with a supported connector or with your own MCP server. If your GPT depended on an external API, plan that part as a separate development and not as a step in the migration.

### Is a project or a plugin better to replace my GPT?

It depends on what you were using it for. If it was a personal or team assistant, made of instructions and documents, a project will resolve it in minutes. If you shared it with many people, want to reuse it in Codex, or need it to communicate with external systems, the correct destination is a plugin.

### How many files can I upload to a ChatGPT project?

A limit count by plan published on October 8, 2026, documents 5 files per project in Free, 25 in Go and Plus, and 40 in Edu, Pro, Business, and Enterprise. This is considerably less than the 20 files that a custom GPT allowed in the paid plan, so in many cases, documents need to be consolidated before migrating.

### Does it affect me if I only use other people's GPTs and haven't created any?

It affects you in access. You do not have to migrate anything —and, in fact, having permission to use a GPT does not enable you to migrate it— but when the creator migrates, access to the old GPT does not automatically grant you access to the new plugin. If you depend on a provider's or a colleague's GPT, ask them now what they plan to do on December 11.

### How do I avoid this happening to me next time?

By keeping the instructions outside of the platform. A document with identity, scope, out of scope, protocol, rules, security barriers, and instruction protection, plus a folder of knowledge files with versioned names, is portable: it can be pasted into a project, converted into a skill, or used in a standalone chat. The platform changes; the document does not.

## The work is in the document, not in the button

What is being retired on December 11, 2026, is a container. The value —the instructions that took you a year to refine, the scope you learned to cut down, the rules that correct the mistakes your assistant made— does not belong to OpenAI: it belongs to you, and fits in a text file. The only truly urgent task in these weeks is to write that file.

And that is where **ChatGPT** fits in, the suite of AI tools from **AiToolsForApps** (not affiliated with OpenAI): its GPT specification generator produces exactly that document —three candidate names, value proposition, system instructions in the seven blocks, four conversation starters, a knowledge plan with the already calculated file names, and a three-phase publication checklist—, adjusted to the limit of 8,000 characters and with the shielding block at the end, which is never trimmed. Its custom instructions module does the same with the two blocks of the account profile, with a character counter and optimizer. You can generate and copy without registering and see the result before deciding, starting with its [prompts by profession and custom instructions for ChatGPT](/chatgpt-prompts-by-profession-and-custom-instructions-229).

There are weeks left. The document can be written in an afternoon.

> Generate your GPT rescue document today: identity, scope, system instructions in seven blocks, four starters, knowledge plan with file names, and a publication checklist. No registration needed to start.

- [ChatGPT — Prompts by profession, custom instructions and custom GPTs](https://aitoolsforapps.com/chatgpt-prompts-by-profession-and-custom-instructions-229)

## About the platform

**ChatGPT — Prompts by profession, custom instructions and custom GPTs** — an app by AiToolsForApps. Available at: https://aitoolsforapps.com/chatgpt-prompts-by-profession-and-custom-instructions-229/blog/custom-gpts-closing-12-december-2026-migrate-plugins-projects

### Other languages

- `es`: https://aitoolsforapps.com/es/chatgpt-prompts-por-oficio-e-instrucciones-personalizadas-229/blog/gpt-personalizados-cierran-11-diciembre-2026-migrar-plugins-proyectos
- `fr`: https://aitoolsforapps.com/fr/chatgpt-prompts-par-metier-et-instructions-personnalisees-229/blog/gpt-personnalises-se-terminent-11-decembre-2026-migrer-plugins-projets
- `it`: https://aitoolsforapps.com/it/chatgpt-prompt-per-professione-e-istruzioni-personalizzate-229/blog/gpt-personalizzati-chiudono-11-dicembre-2026-migrare-plugin-progetti
- `pt`: https://aitoolsforapps.com/pt/chatgpt-prompts-por-profissao-e-instrucoes-personalizadas-229/blog/gpt-personalizados-encerram-11-dezembro-2026-migrar-plugins-projetos
- `de`: https://aitoolsforapps.com/de/chatgpt-prompts-nach-beruf-und-eigene-anweisungen-229/blog/benutzerdefinierte-gpts-schliessen-11-dezember-2026-migrieren-plugins-projekte
- `bg`: https://aitoolsforapps.com/bg/chatgpt-promptove-po-profesiya-personalizirani-instruktsii-i-paketi-po-sektor-229/blog/personalizirani-gpt-priklyuchvat-11-dekemvri-2026-migrira-kam-plagini-proekti
- `nl`: https://aitoolsforapps.com/nl/chatgpt-prompts-per-beroep-aangepaste-instructies-en-pakketten-per-sector-229/blog/aangepaste-gpts-sluiten-11-december-2026-migreren-plugins-projecten
- `ja`: https://aitoolsforapps.com/ja/chatgpt-prompt-builder-custom-instructions-gpts-229/blog/custom-gpt-shuryo-2026-12-11-plugin-project-iko
- `ko`: https://aitoolsforapps.com/ko/chatgpt-prompt-custom-instructions-gpt-229/blog/custom-gpt-retirement-december-11-2026-plugins-projects
- `sv`: https://aitoolsforapps.com/sv/chatgpt-prompter-per-yrke-anpassade-instruktioner-och-branschpaket-229/blog/anpassade-gpt-stanger-11-december-2026-migrera-plugins-projekt
- `da`: https://aitoolsforapps.com/da/chatgpt-prompts-efter-fag-brugerdefinerede-instruktioner-og-branchepakker-229/blog/gpter-lukker-11-december-2026-migrer-til-plugins-eller-projekter
- `no`: https://aitoolsforapps.com/no/chatgpt-prompter-etter-yrke-egendefinerte-instruksjoner-og-bransjepakker-229/blog/egendefinerte-gpter-stenger-11-desember-2026-migrere-plugins-prosjekter
- `id`: https://aitoolsforapps.com/id/chatgpt-prompt-per-profesi-instruksi-kustom-dan-paket-per-sektor-229/blog/gpt-kustom-ditutup-11-desember-2026-migrasi-plugin-proyek
- `cs`: https://aitoolsforapps.com/cs/chatgpt-prompty-podle-povolani-vlastni-instrukce-a-balicky-podle-odvetvi-229/blog/vlastni-gpt-konci-11-prosince-2026-migrace-pluginy-projekty
- `zh`: https://aitoolsforapps.com/zh/chatgpt-prompts-custom-instructions-custom-gpts-229/blog/zidingyi-gpt-guanbi-2026-12-11-qianyi-waigua-zhuanan
- `pl`: https://aitoolsforapps.com/pl/chatgpt-prompty-wedlug-zawodu-spersonalizowane-instrukcje-i-pakiety-branzowe-229/blog/wlasne-gpt-zamykaja-sie-11-grudnia-2026-migracja-pluginy-projekty
- `ar`: https://aitoolsforapps.com/ar/chatgpt-auamr-bhsb-almhna-wtalimat-mkhssa-whzm-alqtaaat-229/blog/gpt-custom-tarhil-ila-plugins-11-december-2026
- `fi`: https://aitoolsforapps.com/fi/chatgpt-ammattikohtaiset-kehotteet-mukautetut-ohjeet-ja-toimialapaketit-229/blog/mukautetut-gpt-t-suljetaan-11-joulukuuta-2026-siirto-plugineihin
- `he`: https://aitoolsforapps.com/he/chatgpt-promptim-lpi-mktsoa-hnchiot-motamot-aishit-vchbilot-lpi-anf-229/blog/gpt-mutamim-ishit-nisgarim-11-detzember-2026-migratzia-plaginim-projektim
