What exactly is changing for custom GPTs?
A custom GPT is a version of ChatGPT set up for one job: instructions, reference files, optional tools such as web search, and sometimes a connection to an outside service. OpenAI is replacing that format with plugins, which bundle "skills" (reusable instructions) and connected apps. The day-to-day difference matters: you no longer open a dedicated assistant. ChatGPT picks a skill when your request matches it, or you call it with @. OpenAI's own FAQ warns that an installed plugin "will not necessarily run on every request".
On personal accounts (Free, Go, Plus, Pro), creating and publishing new GPTs is already closed; existing GPTs keep working until retirement. In Business, Enterprise and Edu workspaces, it depends on workspace settings.
Which dates matter?
The dates below come from OpenAI's retirement and migration FAQ, read on October 5, 2026. OpenAI says they may change and that the notice shown in your account is the one that applies.
| Date | What happens | Who it applies to |
|---|---|---|
| September 11, 2026 | Admin notice to help organisations prepare | Affected Enterprise workspaces |
| September 22, 2026 (target) | Migration experience and in-product banner, with no guarantee every workspace sees it that day | Affected Enterprise workspaces |
| October 26, 2026 (planned) | Creation of new GPTs ends; existing, unmigrated GPTs can still be edited | Affected Enterprise workspaces |
| December 11, 2026 | Retirement: GPTs and their pages become inaccessible | Standard date, all plans (your account notice applies) |
| February 11, 2027 | Deferred retirement | Only Enterprise workspaces with an approved deferral |
Two traps. Several guides that still rank well give September 25, 2026 as the end of GPT creation: that was an earlier version of the FAQ, since replaced by October 26, 2026 (Visual Studio Magazine reported the change on September 28, 2026). And the timeline that applies is the one of the workspace where the GPT was created, not the account you use it from.
How do you take stock of your GPTs?
One row per GPT in a plain spreadsheet. Five columns are enough:
- Name and link of the GPT.
- Creator: only they can migrate it (or a workspace admin on Enterprise). Being allowed to use a GPT does not let you migrate it.
- Who uses it: just you, your team, or outside people (customers, partners, prospects).
- What it contains: instructions and files only, a connected app, or a custom action calling an API.
- Decision: migrate, rebuild, replace with something else, or let it lapse.
Deal with one case first: the GPT built by someone who has left, or by a contractor, on their own account. Nobody else can migrate it. While that person is still reachable, ask for the instruction text and the files. That text is what holds the value: it can be reused in a plugin, a ChatGPT project or another tool.
Before migrating, publish the version you want to keep. Migration uses the latest published version; drafts do not transfer, and publishing only to yourself is enough. Make your last edits first as well: once migrated, the original GPT becomes read-only.
What does migration keep, and what does it leave behind?
| GPT component | After "Migrate to plugin" | What you do |
|---|---|---|
| Instructions | Become a skill in the plugin | Review, then test on your usual requests |
| Knowledge files | Copied into the plugin's reference files | Check they are all there: creators on OpenAI's forum report files rejected during migration |
| Connected apps | Added to the plugin | Each user still needs access to the app and must authorise it |
| Custom actions (API calls) | Not transferred | Rebuild the connection (see below) |
| Selected model | Not carried over | Compare answers before switching |
| Conversations | Not moved | Copy what you need before December 11, 2026 |
| Sharing | Not carried over: the plugin starts private | Share again, and replace outside links |
OpenAI's FAQ says nothing about conversation starters (the sample prompts shown when a GPT opens). Note them in your inventory: they make a ready-made test set for the plugin.
For testing, OpenAI recommends comparing the plugin with the old GPT on familiar prompts and at least one harder case. Same prompts, same files, old GPT in one tab and plugin in the other, while the old one still works.
What should you do with a GPT that calls an API?
This is the only real project. A custom action let the GPT query other software: look up a customer in the CRM, read a product catalogue, open a ticket. It does not go through migration, and anything that depends on it will stop working. OpenAI offers two routes: an existing app that does the same job, or a custom MCP server (a standard connector between an AI assistant and a piece of software), which takes technical setup. The FAQ adds that a rebuilt integration should not be assumed to do everything the old action did.
Before rebuilding, ask one question: does this work really belong in a ChatGPT window? A GPT that reads your CRM to draft a reply is already a small agent, but one you have to open and prompt every time. If the task comes back every day (sorting incoming requests, preparing a quote, following up), it often sits better inside your own tools, with an automation or an agent that works in your inbox and your CRM and brings you anything that commits the business. If the task is occasional, the plugin is enough and you do not need anyone.
Either way, rebuilding gives a new component access to your data. The four questions to ask before connecting an AI agent to your tools apply here: what can it access, who approves what goes out, who can see what it did, and how do you stop it.
How do you replace the links you gave to customers?
None of the six French-language pages we read covers this. Yet some businesses made a GPT available to outsiders: an intake assistant on the website, an "ask us anything" link in a signature, a link in an onboarding document or behind a QR code. Several threads on OpenAI's forum raise it, one titled "how do small businesses keep serving external customers after Dec 11?" (listed by Visual Studio Magazine on September 28, 2026). On December 11, 2026, those links will lead to an inaccessible page.
The plugin does not take over on its own. According to OpenAI's FAQ, it starts private, people who used the old GPT do not get access automatically, using it requires an account where plugins are available, and making it public goes through a separate submission process whose timing the FAQ does not give. A customer without a ChatGPT account will have no replacement to click.
- Find the links. GPT addresses start with chatgpt.com/g/. Search your website, email and signature templates, sales documents and printed material.
- Decide, link by link. Keep an equivalent service (a form, a FAQ page, an assistant hosted on your own site), or remove it.
- Replace before retirement, not on December 10, 2026: a printed leaflet or a signature already sent cannot be corrected.
A GPT cannot be embedded in a website: OpenAI says so in its FAQ and points to its API for building an assistant into a product. If your GPT was the first answer your customers got, this is the moment to decide where that answer should live and who checks it.
What timeline should a small business follow?
- This week: the inventory, and recovering the instructions of GPTs created by someone else.
- Before October 26, 2026, if you are on an Enterprise workspace: publish any drafts you want to migrate.
- October and November: migrate and test the simple GPTs; start rebuilding the ones that call an API, which takes longest.
- End of November at the latest: share the plugins, have one user test them, replace outside links.
- December 11, 2026: GPTs stop. Anything not copied is gone.
Our own agents have run since September on instructions written in skill files, the format OpenAI is moving GPTs to. What we take from it: what lasts is the text of the instructions and of the documents that are the source of truth, not the tool that reads them. A GPT whose instructions are cleanly written changes format without damage; a GPT tuned as you went, with no written record, is the one that will break.
FAQ
Will my GPTs stop working before December 11, 2026?
No. According to OpenAI's FAQ, existing GPTs remain usable until the retirement date, including after migration, in read-only mode. Only Enterprise workspaces with an approved deferral run until February 11, 2027.
I don't see "Migrate to plugin". Is that normal?
Often, yes: migration does not reach every account on the same day, and plugins may be disabled in your workspace. Check that you are signed in to the account where the GPT was created and that you are its creator; on Enterprise, a workspace admin can also migrate it.
Will the plugin answer like my GPT?
Not necessarily: OpenAI writes that a migrated plugin may respond differently, partly because the selected model is not carried over. Test it on your usual prompts and one hard case, side by side with the old GPT, before sharing it.
I use a GPT someone else built. What should I do?
Nothing on the migration side, which you cannot do yourself. Watch the notice on the GPT and its creator's guidance, and check that you actually have access to the replacement before December 11, 2026.
Sources: "Custom GPT retirement and migration FAQ", OpenAI Help Center, read October 5, 2026 (help.openai.com) · "GPTs in ChatGPT", OpenAI Help Center, read October 5, 2026 (help.openai.com) · D. Ramel, "Hands-On With OpenAI's Mandatory Custom GPT-to-Plugin Migration", Visual Studio Magazine, September 28, 2026, for the move from September 25, 2026 to October 26, 2026 and creators' reports (visualstudiomagazine.com) · Équipage IA review of October 5, 2026: Google France organic results for "fin des gpt personnalisés", six French-language pages read in full (excluding OpenAI's FAQ and one unrelated article); two still give September 25, 2026 as the end of creation, none covers GPT links placed on a website, a signature or a customer document.