Custom GPT Retirement: What to Do Before Dec 11

On September 11, OpenAI put a line in its ChatGPT release notes that a lot of small businesses will only notice in December: custom GPTs are being retired. The scheduled date is December 11, 2026. If you built a custom GPT for your team, yo

On September 11, OpenAI put a line in its ChatGPT release notes that a lot of small businesses will only notice in December: custom GPTs are being retired. The scheduled date is December 11, 2026.

If you built a custom GPT for your team, your customers, or your own sanity, you have a to-do list now. The custom GPT retirement is not as bad as the headlines make it sound for most GPTs. For one type of GPT, it's a rebuild.

What OpenAI has confirmed so far

We read the primary sources instead of the recaps. OpenAI's custom GPT retirement and migration FAQ says the transition affects all ChatGPT plans, that existing GPTs stay usable until the retirement date, and that creators get a Migrate to plugin button under My GPTs once migration reaches their account.

Two caveats matter. Migration is rolling out account by account, so a missing button today doesn't mean you did something wrong. And OpenAI's own wording is "scheduled" and "planned," with Enterprise dates flagged as subject to change. Treat December 11 as real and plan backward from it.

The replacement is a plugin: a package of reusable instructions, reference files, and connected apps, described in OpenAI's plugins documentation.

What carries over and what dies

Sort your GPTs by what they contain. Here is the migration, straight from OpenAI's FAQ:

• Instructions become a skill inside the plugin.
• Knowledge files are copied into the plugin's reference files.
• Connected apps are added to the plugin as apps.
• The model you picked does not carry over.
• Existing conversations do not move.
• Sharing settings do not move, and the plugin starts private.
• Custom actions do not transfer at all.

GPT custom actions do not transfer through the migration workflow. Features that depend on them will not work in the plugin until you set up a replacement.

— OpenAI, Custom GPT retirement and migration FAQ

That last line is where the real work hides. A GPT with instructions and a PDF or two is a prompt with a URL. It migrates in minutes. A GPT with a custom action, the kind that writes a row into your CRM, checks a booking calendar, or fires a Zapier zap, is a small piece of software. OpenAI says rebuilding that connection may require a custom MCP server and technical setup, and warns that a rebuilt integration shouldn't be assumed to do everything the original did.

The one hands-on test we found

We haven't migrated a client GPT ourselves yet, so we won't pretend to. The one practical test we could find is from The Paperless Agent, which migrated a real GPT that imports contacts into a CRM through a Zapier action. ChatGPT warned that the custom action wouldn't work in the plugin. The instructions moved over completely. The Zapier connection did not.

That's a single test on a single Plus account, so treat it as a data point. It matches what OpenAI's FAQ predicts, which is the useful part: the docs and the real behavior agree.

Sort every GPT into one of four buckets

Before December, list every custom GPT that your business touches. Include the ones your assistant, your contractor, or that one employee who "just made a quick GPT" set up on a personal account. Then sort.

Bucket 1: instructions only. A brand-voice writer, a proposal template, an SOP explainer. Migrate it, test it against three prompts you know the answer to, and move on. This is a 20-minute job.

Bucket 2: instructions plus reference files. An employee handbook bot, a product-spec assistant, an FAQ answerer. Same migration, but read the converted reference files. OpenAI itself says a migrated plugin may respond differently, so compare answers on your hardest questions, not your easiest.

Bucket 3: GPTs with a custom action. Anything that reads or writes to another system. Don't migrate these on autopilot. Ask first whether a built-in ChatGPT app already reaches that tool. If it does, you may not need to rebuild anything. If it doesn't, this is a workflow project.

Bucket 4: GPTs other people use. A lead-magnet GPT on your website, a customer-facing assistant, a tool you share with clients. Access to the old GPT does not give anyone access to the new plugin. Public sharing doesn't carry over, and a public plugin needs its own submission process. Your customers will hit a dead link on the retirement date unless you've told them where to go.

Bucket 3 is where the real decision sits

Our view: a custom GPT was never the system. It was the front door. The system was the action behind it, and you were renting the front door from someone who just told you the locks are changing.

Rebuilding that action inside OpenAI's plugin format means learning the new format and hoping it holds still. In June, OpenAI retired GPT-4.5 in ChatGPT, custom GPTs included. Now it's retiring the whole format. Two forced changes in one year is a pattern, not bad luck.

Our opinion: put the logic where you own it. A workflow in n8n that catches a lead, writes to the CRM, checks the calendar, and asks a model to draft the reply treats the model as one swappable step, not the foundation. When a vendor changes the deal next time, you change one node instead of rebuilding the process. We laid out where that trade lands in n8n vs Claude agents: when each wins.

It also fixes a problem GPTs never solved: who's running what. Half the GPTs in a typical small business live on one person's personal account, which is the same exposure we flagged in what happens when your team pastes client data into ChatGPT. A workflow you own has a login, a log, and an owner.

Test the replacement like you'd test a new hire

A migrated plugin isn't a copy. It's a translation, and translations drift. OpenAI's FAQ lists four things to check, and they make a good acceptance test:

• Does it pick the right skill and follow your instructions?
• Does it use the reference material you expected, not a plausible guess?
• Does it produce complete answers or files in the format you need?
• Does it have the tools and integrations the workflow depends on?

Write down five real prompts before you migrate. Use three routine ones, one that should trigger a refusal or a handoff to a human, and one ugly edge case from last month. Run the original GPT and the plugin side by side and keep the transcripts. If the plugin fumbles the edge case, that's your answer on whether it's ready.

The refusal prompt matters most. A GPT that used to say "I can't quote that, please call the office" and now cheerfully invents a price is the one that costs you money. Nothing in the migration guarantees the old guardrails survived the translation.

What a plugin still won't do for you

Plugins are a real improvement over GPTs for bucket 1 and 2 work. The plugins documentation says apps inside a plugin stay subject to their own authorization and action-approval rules, and plugin creation on the web is described for Business and Enterprise workspaces. If you're on a personal plan, check what your account actually offers before you promise a colleague a shiny replacement.

The bigger limit is the one we keep coming back to. A plugin waits for a person to open ChatGPT and ask. As we wrote about ChatGPT Work, it's an excellent assistant and a poor substitute for something that runs when a lead arrives at 11pm with nobody at a keyboard. And voice mode now taps the same plugins, so the approval rules for anything consequential still apply there too.

If your custom action existed because you wanted something to happen without you, a plugin is a step sideways, not forward.

A dated plan for the next 73 days

Roughly in order:

1. This week. Inventory every GPT. Note who created it, which account it lives on, and whether it has an action.
2. By mid-October. Migrate buckets 1 and 2. Test each against your known-answer prompts. Publish any drafts you still need before migrating, since unpublished edits don't move.
3. By early November. Decide bucket 3: rebuild inside ChatGPT, or move the logic into a workflow you own. Don't decide in December.
4. By late November. Tell everyone in bucket 4 where to go, and confirm they can actually open the replacement.
5. Before December 11. Leave the read-only originals alone until you've watched the replacements run for a week.

Enterprise workspaces have their own timeline, including a planned end to creating new custom GPTs on October 26 and an approved-deferral retirement date of February 11, 2027. If that's you, follow your admin notice, not this list.

If bucket 3 has more than one GPT in it, that's a project. It's the kind we build: the instructions, the CRM and calendar connections, and the approval step, as one automation you own, finished ahead of the date.