Blog AI strategy
Custom GPTs Are Becoming Plugins and Skills: What Actually Changes
OpenAI is retiring Custom GPTs on 11 December 2026. The deadline is the least interesting part. What changes is the unit of reusable AI: from an assistant you visit to a capability that comes to you.
For three years the easiest way to build something useful on ChatGPT was a Custom GPT. Give it instructions, upload some knowledge, connect an action, share it with the team. No application to build.
OpenAI is retiring that. Custom GPTs stop running on 11 December 2026, and the path forward is Plugins and Skills. Admin notices went out on 11 September 2026, and the milestones in between differ by plan and by workspace, so the dates that matter to you are the ones in your own notice rather than any published elsewhere.
The deadline is the least interesting part of this. What changes is the unit of reusable AI.
From building an assistant to building a capability
A Custom GPT was a destination. You made Marketing GPT, or Contract Review GPT, and people went there to use it. Every GPT was its own island, which is why most teams ended up with five of them and used two.
A Skill is not a destination. It is workflow knowledge that ChatGPT can bring into whatever conversation the user is already having.
| Custom GPT | Plugin and Skill |
|---|---|
| Build a specialised chatbot | Build a reusable capability |
| The user opens your GPT | The capability comes to the user’s conversation |
| Instructions belong to one GPT | Skills package workflow guidance anything can use |
| Actions connect that GPT to an API | Apps provide access to external systems |
| Think assistant | Think workflow |
Skill, App, Plugin: the short version
- A Skill is how. Instructions, rules, examples, reference files and sometimes code, describing how to do a job properly.
- An App is what it can reach. The connection to Drive, Slack or another system, inside the permissions the user already has.
- A Plugin is the package. Skills, Apps, or both, shipped together.
An onboarding plugin might hold one Skill describing your onboarding checklist and the tone you write policies in, plus a Drive App that can find the handbook. That is a packaged workflow, not another chatbot.
What carries over, and what you rebuild
OpenAI’s migration takes a published GPT and produces a Plugin. Not everything makes the trip.
| Part of your GPT | What happens |
|---|---|
| Instructions | Become the Skill |
| Knowledge files | Become reference files inside it |
| Connected apps | Added to the Plugin |
| Custom actions | Do not transfer. Rebuild them as a connector or your own server |
| Selected model | Does not carry over |
| The original GPT | Stays usable until retirement, but becomes read-only |
That last row has a practical consequence: finish any edits you care about before you migrate, because afterwards the old one is frozen. It also gives you a parallel-running window, which is the only sane way to do this.
The part worth protecting is not the prompt
A GPT that has been in daily use for six months is rarely just a prompt. It holds decision rules, edge cases someone hit in week three, formatting nobody wrote down, and an order of operations that took a few arguments to settle.
Tina Austin makes this point from a teaching perspective in Making the move from Custom GPTs: educators poured real teaching judgement into these tools, and a migration button does not carry judgement across. It applies just as well to a support team’s triage GPT.
Before you migrate anything, write down six things per GPT:
- The instructions, and which parts are actually load-bearing
- The knowledge it depends on
- The tools and actions it calls
- The workflow logic, meaning the order it does things in
- Who relies on it today
- Five prompts whose answers you can compare before and after
The last one is the one people skip. Migration should read: known behaviour, test cases, migrate, run the same tests, compare, fix, deploy. Treat it like a software migration, because it is one.
Your instructions are more portable than they look
Here is the thing worth noticing while everyone is busy with deadlines. A Custom GPT’s instructions and knowledge files are not locked to ChatGPT. A Claude Skill has the same shape: a markdown file of instructions, plus reference files it loads when the task needs them.
So the same inventory that prepares you for OpenAI’s migration also lets you take that work anywhere. We built GPT2Skill to do the second half: paste a Custom GPT’s name, description and instructions, drop in its knowledge files, and get a Claude Skill as a zip you can upload. It runs entirely in your browser, so the instructions you paste never reach a server, and it is free.
It is not a replacement for OpenAI’s migration, and it does not pretend to carry actions across either. It is for the case where you would rather your capability was not tied to one vendor’s roadmap.
What we would do this month
- List every GPT anyone actually uses. Most teams find half are dead.
- For the survivors, write the six-item inventory above.
- Migrate one, run both side by side for a week, compare on your test prompts.
- Rebuild actions deliberately. An API integration is a connector with its own permissions, not a paragraph of instructions.
- Then do the rest, and keep the inventory. It outlives whichever platform.
The bigger shift
First we had chatbots, then specialised chatbots, then agents. Now the pieces are becoming components that an agent assembles for the task in front of it. A Skill does not need a personality. An App does not need a chat window. The conversation is the centre, and capabilities come to it.
That is why the interesting headline is not that OpenAI is killing Custom GPTs. It is that the unit of reusable AI just changed from the chatbot to the capability. The question to ask of each of your GPTs is not how to recreate it, but which part of it was reusable intelligence, which part was access to data, and whether either is useful well beyond the GPT it was trapped in.
Working through this? GPT2Skill converts a Custom GPT into a Claude Skill in your browser, free and with nothing uploaded. If the answer is really an agent rather than a skill, four of ours run free on this site.
Open GPT2SkillSources
- OpenAI, Custom GPT retirement and migration FAQ (the authority on your own workspace’s dates)
- Mixed News, OpenAI is retiring custom GPTs: what carries over to plugins
- Tina Austin, Making the move from Custom GPTs to Plugins, 22 September 2026
Dates are OpenAI’s published plan and may change. Intermediate milestones vary by plan and workspace, and reports of them differ, so treat the notice in your own workspace as the one that counts.