On September 11, OpenAI announced that the custom GPTs in ChatGPT are going away. A GPT is your own version of ChatGPT: you give it fixed instructions and your own files, and from then on it does one kind of job. Drafting a quote in your house style, answering a customer's question based on your terms and conditions, summarising a report the way your team likes it. OpenAI introduced them in November 2023, and many businesses have put real work on them since.
According to the explanation OpenAI itself published, you can no longer create new GPTs from October 26, and December 11 is the standard date on which they stop. Business workspaces with an approved deferral keep them until February 11, 2027. The change applies to all plans, although the timing may differ per account. The first reports, in mid-September, still gave September 25 as the last day to create a GPT. The calendar belongs to the provider, and it moves.
The successor is called a plugin: a package that brings together instructions, files and connections to other software. OpenAI offers a button to switch over. That sounds like a move where everything gets packed for you. Read the list of what comes along and what doesn't, and you see something else. In it, I think, lies the lesson for every business that uses AI, whichever provider it relies on: your AI assistant is rented, and the only part that is yours is what you have written down and tested outside the package.
What comes along and what stays behind
The instructions you wrote become a so-called skill inside the plugin. Your files are copied over. Apps you had connected through ChatGPT itself come along too. That is the part made of text and choices you recorded yourself.
What stays behind, OpenAI lists itself:
- Custom connections to your own systems, such as a GPT that looks up a customer in your planning tool or records an order. According to OpenAI, rebuilding them may take technical work.
- The AI model you had chosen, the part that actually writes the answers.
- Your existing conversations.
- Access. A migrated plugin starts out private, and the people who used your GPT do not automatically get access to its replacement.
Then comes a sentence I would like to read aloud to every business owner: a migrated plugin may respond differently. OpenAI advises asking familiar questions again, plus at least one harder case, and checking whether it still follows your instructions.
Put those two lists side by side. What comes along is what you wrote down yourself. What stays behind is everything that made the assistant work inside your business: the connections, the people who used it and the behaviour you relied on.
The biggest risk is the one you don't see
A GPT that stops on December 11, you notice on December 11. Annoying, but visible. Harder is an assistant that keeps working after the move, just slightly differently. A different model with the same instructions gives different answers. Say your assistant prepares quotes and after the switch it misses a discount rule a little more often, or addresses a client too casually. Nobody notices right away, because nobody ever wrote down what a good answer looks like.
On OpenAI's own forum you can read what this feels like in practice. A small business in Brazil writes that it supplies 47 GPTs to paying customers, shared through a link. That way of sharing does not carry over. Other users report that a migrated plugin used up far more of their usage allowance, and that it took a lot of trying to get it to do what the original did. These are individual experiences, and I would not draw conclusions about cost from them. But they show where it hurts: in behaviour and access, exactly what doesn't fit in a moving box.
That is why I consider OpenAI's testing advice the most important line in this whole story. Asking familiar questions again only works if you have those questions, with the answers you expect, written down somewhere. A fixed set of test questions turns “it seems to be going fine” into something you can check.
This is how I work myself. Gemeenteregels answers your question about what is allowed in your municipality with the article from the official rules, and for advisers it also puts municipalities side by side. To do that, the system has to pick the right article in each municipality. I test that with a fixed series of questions whose correct answers I know. Just before the site went live, that series showed it got it right in 76 percent of cases. After a round of improvements, it was 97 percent. Without that series, all I could have said was that it felt better. And since September 27, a weekly check compares the stored search data with the official text and repairs anything that deviates.
The strongest counterargument
You can build a GPT in an afternoon. If it disappears, you rebuild it in an afternoon. Why put effort into something cheap and temporary? And anyone who starts building everything themselves is stuck with maintenance, while the big providers improve faster than you can keep up.
Partly, that is true. An assistant that only rewrites emails does not need protecting: rebuild it after the switch and move on. And building your own does not have to be the goal. I work with other companies' AI models every day myself.
But that afternoon is the first version. An assistant that real work runs on has usually been refined for months after that: a rule added after every wrong answer, an exception for that one client, a file that was replaced. That knowledge often lives only in the package, and sometimes only the colleague who built it knows why a rule is there. Starting over then means losing those months, or having to retrieve them from the memory of someone who may already have left.
So my answer to the counterargument is modest. Feel free to use a big provider's package. Just make sure three things also exist outside it: your instructions and files, your test questions with the expected answers, and a description of the connections to your own systems. Then your choice of provider becomes something you can reverse.
What you decide before December 11
If you have GPTs that real work runs on, use the move to detach them from the package. Concretely:
- List every GPT used in your business, who uses it, inside and outside the company, and what goes wrong if it answers differently for a week.
- Copy the instructions and files into your own documents, written so that a new colleague would understand them too.
- For each assistant, write down ten questions with the answer you expect, including a few hard ones. Ask them before and after the switch, and after every change from then on.
- Find out which assistants have a custom connection to your systems. That is where most of the work is, so that is where you start.
You can treat this as a chore that is done by December 11. Or you make sure now that your instructions, test questions and connections live in a place of your own, so that with the next change, from whichever provider, you know within a day whether your AI still does what you expect of it. Only the second choice leaves you with something that stays yours.
Sources
- Custom GPT retirement and migration FAQ, OpenAI Help Center, 2026-09-11
- OpenAI is retiring custom GPTs, here's what carries over to plugins, MIXED, 2026-09-18
- Custom GPTs in ChatGPT are going away. Here's how to save yours, PCWorld, 2026-09-18
- OpenAI's Custom GPT to plugin migration raises access concerns, OpenAI Developer Community, 2026-09-23