OpenAI's Custom GPT Migration Won't Carry Actions to Plugins

If a custom GPT in your workspace calls an external API, the one-click migration will not keep that workflow running, and you have 40 days to rebuild it. Custom GPTs are scheduled to become inaccessible on December 11, 2026, according to Developers Digest's summary of OpenAI's retirement FAQ. OpenAI's migration guide says plainly that "custom actions do not transfer through the migration workflow." Instructions, knowledge files and connected apps do move. So the job this month is to find the GPTs with actions, decide which ones deserve a rebuild, and give your engineers enough lead time to do it.

A custom action is a connection that lets a GPT in ChatGPT call an outside API: the model picks the call, generates the JSON input from a schema you described, and the action executes it. A plugin is OpenAI's replacement package, which can hold skills (instructions for a task), apps (connections to other services), or both.

What Moves to a Plugin and What Stays Behind

Text and files move, while integrations and access settings mostly do not. This table combines OpenAI's migration guide with Developers Digest's reading of the FAQ:

GPT feature	After migration	What you do
Instructions	Become a skill	Test that the skill triggers; use an @ mention where it must run
Knowledge files	Copied as reference files	Check that answers still cite the right material
Connected apps	Stay in the plugin	Confirm each app is enabled and accounts have access
Custom actions	Do not transfer	Rebuild on a supported app or a custom MCP server
Selected model	Does not carry over; Enterprise defaults apply	Rerun familiar prompts and compare the output
Sharing	Replacement starts private	Share again, then have a recipient install and test it
Drafts	Only the latest published version moves	Publish needed edits first

Migration also freezes the original. The GPT stays usable until retirement but turns read-only, and its creator cannot delete it. Past chats do not move to the plugin, per EdTech Innovation Hub. OpenAI's FAQ, as excerpted in search results on November 1, 2026, says existing conversations stay accessible after retirement with no action needed. The migration guide still tells you to save useful prompts and identify the context you need.

The Dates, and Where the Sources Disagree

December 11 is the date to plan around, and OpenAI itself says it can move. The FAQ's wording, as quoted by Developers Digest, is that dates are "subject to change" and "a target date is not a guarantee." Virtualization Review adds that timing "can vary by account and workspace." OpenAI's Help Center returned a 403 to a scripted fetch on November 1, 2026, so the dates below come from write-ups that cite it, and the FedRAMP details come from its GPT help article as excerpted in search results that day. Check the live FAQ from a browser before you send a date to your organization.

Date	Milestone	Scope
September 11, 2026	Admin notice	Enterprise
October 26, 2026	New custom GPT creation ended; unmigrated GPTs stay editable	Affected Enterprise workspaces except FedRAMP, per Developers Digest, Virtualization Review and OpenAI's help article
December 11, 2026	Standard retirement	All plans
February 11, 2027	Retirement for FedRAMP workspaces and workspaces with an approved deferral	Some Enterprise

Two gaps remain. iTechnically Kan lists the October 26 cutoff without limiting it to Enterprise. The FAQ excerpts from that day say personal accounts (Free, Go, Plus and Pro) cannot create or publish new GPTs, and they send Business workspaces to their workspace notice and settings, so Business owners should check there. And neither Virtualization Review nor EdTech Innovation Hub explains what qualifies for a deferral or how to request one. Ask your OpenAI account team in writing, and keep planning for December 11 until you hold an approval. The same excerpts say a deferral alone does not preserve the ability to create new GPTs. FedRAMP workspaces are listed for February 11 and can still create GPTs, so confirm that date with the same team.

---

Why Custom Actions Are the Expensive Part

A rebuilt action is a second plugin with its own permissions, hosting and testing. OpenAI's rebuild guide has you ask the original GPT for a rebuild prompt, check that prompt against your saved action settings, and create a new Sites plugin, where Sites hosts the MCP server. The result is separate from the migrated plugin, and you can use both in one conversation. Sites plugins do not currently support editing skills or reference files. Plan on one old GPT becoming two things to own.

The permissions catch is the part most likely to surprise an admin. The migration guide says built-in migration does not need Upload plugins or Upload plugins with custom MCP servers. The rebuild guide says the Site owner publishing an MCP-backed plugin needs Upload plugins and Create plugins with MCPs, both off by default in Enterprise. So a workspace can finish migration prep and still block every MCP rebuild. The two pages also name that second permission differently, so check the label in your own admin console.

Access changes as well. Teammates need at least viewer access to the plugin, its app and the Site, and each person "installs and connects with their own account." If your old action ran on one shared service credential, compare the two models with your security lead, because per-user sign-in can change who reaches which records.

The case for the rebuild is real. OpenAI's MCP guidance tells builders to "enforce authorization in the MCP server for every request; never rely on the model to decide whether a user has access," and to set readOnlyHint and destructiveHint to match what each tool does. The same plugin package and public directory serve ChatGPT and Codex. An action that sat inside one GPT becomes a server with an owner, which your security team can review once and which other plugins can reuse. For the controls around that server, see our MCP server governance guide, and test the tool design: a generic MCP connector lost to no tools at all in UNICEF's benchmark.

How to Host the Rebuild When the API Is Private

Your hosting choice comes down to where the API lives. The default in OpenAI's rebuild guide is a Sites plugin that reuses your existing API where possible, and the server still needs an approved connection to that API. For another host, you follow OpenAI's Build an MCP server guide, which calls for a streamable HTTP endpoint and OAuth 2.1 for user login.

For systems behind your firewall, Secure MCP Tunnel connects private servers without inbound firewall rules: a customer-run tunnel-client polls OpenAI for queued requests and forwards them. It needs outbound HTTPS to api.openai.com:443 and does not support public plugin distribution. Two details deserve a test before you commit. A tunnel tied only to a personal Platform organization "doesn't automatically appear in an Enterprise/Edu workspace," and tunnel permissions sit at the Platform organization level while ChatGPT access is governed by workspace permissions. The page does not describe a general workspace rollout, so prove it with one real teammate account first.

Auth secrets stay out of the package. Plugin manifests declare connection settings, not credentials, and the rebuild guide says to leave credentials out of the prompt you hand to Plugin Creator.

---

What Migration Costs

Conversion is free, and the running costs depend on your pricing model. OpenAI's guide says converting a GPT to a plugin does not consume credits, while editing or updating one does under credit-based pricing. It adds that Instant Chat is "generally unlimited" under that pricing but a request can switch to a reasoning model, and that Work consumes credits based on model and tokens. iTechnically Kan cites the FAQ as saying that, in Enterprise workspaces, a migrated GPT without apps in Instant Chat has no incremental cost.

So a simple prompt GPT may cost nothing more, while a workflow that routes through Work now draws on your credit pool. OpenAI also notes that a zero overage limit prevents spending beyond the shared allocation. Run your top five replacements for a week and read the usage before you roll them out. Our earlier piece on ChatGPT Enterprise spend controls covers the admin side.

This Is the Second Time OpenAI Has Done This

ChatGPT builders have been through a forced move before, and GPTs were the official replacement then. When OpenAI wound down the original plugins beta, Zapier told customers that new conversations would end on March 19, 2024 and existing ones would stop working on April 9, 2024. Zapier now points those users to Zapier MCP.

Google is running a similar change on a slower clock with a kinder default. Its admin guide says Business and Enterprise users lose Gems no sooner than March 1, 2027, and that Drive permissions and sharing settings carry over to the converted skills, which arrive as inactive drafts. Gems used in Workspace apps have to be recreated by hand. Our Gems coverage has the details. Skills are also moving toward a shared format: Codex and GitHub Copilot read Anthropic's Agent Skills, which makes instructions the most portable thing you own here.

A correction to our own record: our April story on Workspace Agents replacing custom GPTs carried a May 15 date. Current reporting puts the retirement at December 11.

---

The Runbook for the Next 40 Days

This Week:

1. Run the GPT report. In ChatGPT Work, type @ and select ChatGPT Admin, then ask for recently used GPTs, which lists owners, usage, migration status and custom actions. Export it to a sheet.
2. Add what the report can miss. OpenAI's suggested prompt prioritizes the last 30 days, so ask finance, legal and audit owners for GPTs they run quarterly.
3. Name one owner per GPT, because only the creator or a workspace owner or admin can migrate. Access to use someone else's GPT does not grant that right.
4. Send your account team the deferral question in writing, plus the FedRAMP date if it applies to you.

By November 14:

1. Sort every GPT into one of three piles: migrate as is (no actions), rebuild (actions that a live workflow depends on), or retire (no recent use or no owner).
2. Check the Plugin Directory for an existing app before you build anything for a SaaS service.
3. Turn on Enable plugins and Use plugins, then grant the MCP permissions to a named group of builders. Save each action's schema, auth method and non-secret settings before you migrate.
4. Publish needed edits. Never-published drafts are the open risk after the October 26 cutoff, so have owners try this now and escalate if it fails.

By December 1:

1. Finish final edits, then migrate in batches, since originals become read-only. Bulk migration does not copy sharing, and OpenAI says only that it expects bulk sharing "soon," so budget for manual sharing.
2. Have owners test with familiar tasks and one harder case, using an account with the same permissions as a typical teammate. Skills are not always triggered automatically, so require an @ mention on any workflow that must run. OpenAI advises testing over assuming identical behavior.
3. Have security review each rebuilt server's authorization and tool annotations.
4. Update onboarding docs and internal links that point at old GPT pages.

Before December 11: Switch users over, and keep a list of anything still unrebuilt along with who has accepted the outage. OpenAI's guide says migration stays possible after retirement, though the GPT will not run, so instructions and files are recoverable while a dependent workflow is down.

The Bottom Line

A GPT without actions takes one click and a test. A GPT that calls your systems needs engineering time, a permission that is off by default and a security review. Budget for the second kind as an integration project with a hard date, and use the inventory to retire the GPTs nobody owns.

Run the GPT report today and count the rows with custom actions.

Continue Reading

• OpenAI Keeps September GPT-6 in Work and Codex as Chat Gets October
• Google ends Gemini Gems for Business and Enterprise in 2027
• MCP Server Governance: Enforce at the Client, Not the Registry
• UNICEF Tested a Generic MCP Server. It Lost to No Tools.
• LibreChat and Open WebUI offer ChatGPT Enterprise alternatives
• GPT Spend Management: OpenAI Enterprise Controls Explained