Is Your Custom GPT Ready to Migrate? A One-GPT Readiness Check for ChatGPT Enterprise

Last updated: September 21, 2026

OpenAI plans to retire custom GPTs on Dec 11, 2026, with a migration path to plugins. In an Enterprise workspace, the planned migration turns a GPT’s instructions into a plugin skill, copies its knowledge files into reference files, and adds its connected apps as apps. Custom actions, conversations, sharing settings and the selected model don’t carry over. This briefing is for whoever owns one GPT a team relies on and must decide its next step.

Private does not mean reversible. OpenAI says migration makes the original GPT read-only and that its creator cannot delete it, and its FAQ describes no way to undo a migration, so the evaluation itself is a one-way step. The second caveat is timing: OpenAI’s own pages give two different migration timelines (section 01), so plan from the notice in your workspace. FSR has not run a migration; everything here comes from OpenAI’s published documentation.

01 What is changing, and which dates apply

OpenAI’s migration FAQ says custom GPTs “are scheduled to retire on Dec 11, 2026,” when they stop running and leave the GPT directory. Until then, existing GPTs stay usable, subject to their existing access and workspace permissions. The same page tells readers to follow “the notice for your account or workspace, including any different timeline communicated to you.”

For the steps before retirement, OpenAI’s pages disagree. This is how the two main pages read on 20 September 2026:

A notice at the top of OpenAI’s page on managing GPT access also targets Sep 17, 2026 for migration, names Dec 11 for retirement, and gives no creation cutoff. Every page hedges: the FAQ says “the dates are subject to change,” the release notes say “Dates may change,” and the GPT access page says migration may not reach every account or workspace at the same time.

The French edition of the same FAQ, marked as machine-translated from English, gives the Sep 17 and Sep 25 dates. The pages don’t show which version came later, so this briefing calls neither date a postponement or the latest word.

Four statements in the English FAQ decide how to read that table. A target date “does not guarantee that every account can migrate on that day.” Until migration reaches your workspace, a missing option “does not by itself indicate a problem.” With several accounts or workspaces, the timeline and permissions of the one where the GPT was created apply. And your admin or account team “can help confirm the timeline that applies.” Plan from your notice, not from this table.

One date needs action before the others. Migration uses the GPT’s latest published version, and the FAQ tells creators to publish “any drafts you need to migrate before this cutoff,” meaning the end of new GPT creation. Publishing doesn’t require sharing the GPT publicly. The cutoff doesn’t retire existing GPTs; they keep running until retirement. Your notice sets the date that applies to you. Since Sep 25 and Oct 26 are both in circulation, finish and publish the edits you need migrated well before the earlier one rather than counting on the later one, and don’t push unreviewed changes into a live GPT to beat a date.

02 What moves, what doesn’t, and what OpenAI doesn’t say

Migration copies a published GPT into a new plugin that starts private. The FAQ describes this mapping:

The original changes too: after migration it “remains usable until retirement, but becomes read-only after migration and its creator cannot delete it,” and future changes go into the plugin.

Several things a GPT owner would want to know are not documented in the pages FSR read:

• whether a migration can be undone or repeated, or the original restored
• what file types, counts or sizes the plugin’s reference files accept (the GPT side has documented limits, 20 files of up to 512 MB each, and the pages don’t say whether the same limits apply after migration)
• how built-in capabilities such as web search, image generation, canvas or code interpreter carry over
• what happens to GPT conversations and files after retirement
• conversation starters, in the English FAQ

Treat each as unknown rather than as a no. Before you migrate, make sure the records you would need to rebuild the GPT exist where your organization keeps them: the published instructions, the list of knowledge files and the conversation starters. Follow your own retention and access rules; copying regulated content into a personal store to make a backup creates a different problem.

If your organization needs past GPT conversations for its records, OpenAI’s GPT access page says conversations with GPTs are available in the Compliance Platform. The migration FAQ doesn’t say whether retirement changes that.

03 The readiness check: six gates for one GPT

The check below is FSR’s, built from OpenAI’s documentation. It isn’t an OpenAI status or tool. For one GPT, it answers one question: start a private evaluation now, fix something first, wait on one item, or retire the use? Evaluation here means migrating and testing the plugin while it stays private. The moment you share it with anyone, even one colleague, you have started a pilot, which is a separate decision (section 08).

One GPT can sit at B on one gate and C on another. Keep both reasons rather than collapsing them into a single letter, and keep preparing everything else while you wait.

Reading the result: if G0 says D, stop there. A needs every gate inside the evaluation’s scope settled, not just migration being available. A B inside that scope means fix first, with any C confirmations running in parallel; a C with no B means wait on that item, with a date. An unanswered question is C only when it blocks this decision: a notice that hasn’t arrived stops the timeline gate, not the work on integrations, people or data. A capability stuck at B stays out of the evaluation until it’s fixed.

And remember what A costs: the evaluation is a real migration, with no undo described, so finish and publish essential edits first.

04 Integrations: connected apps vs. custom actions

Start by naming the mechanism. OpenAI’s GPT configuration page is explicit: “A GPT can use either apps or actions, but not both at the same time.” So your GPT sits on one side of the table below, and that side decides whether the integration moves at all. Then ask what the connection can do, which decides how carefully you test it.

A connected app keeps its own gates after migration. The FAQ is explicit: “Installing a plugin or connecting an account does not give you additional permissions to files or data in the connected service.” In the admin settings, role access decides who can use an app, action settings decide what it can do, and the approval setting decides when ChatGPT asks first. Plugins that include required apps inherit those apps’ role and action settings, and some changes “may require a user to reconnect or reauthorize the app.” Check how the app actually authorizes before you plan the rollout: OpenAI describes setups that use an individual account, a no-auth app, or an administrator-managed connection, so not every app means a personal sign-in for every user.

Don’t assume the apps are on for the people who will use the plugin: OpenAI says new Enterprise workspaces start with a selected set of apps enabled, that those defaults “do not change existing workspace settings,” and that in general new plugins and apps are disabled by default. Your admin has the answer for your workspace.

Then sort each connection by what it can do. Reading data, changing data, and actions that are hard to undo need different tests. Apps can do more than read: OpenAI describes plugins using a connected app to “search information, retrieve content, or complete supported actions,” and admins enable read and write actions separately. Before you test anything that writes, agree on the test data, an authorized target you control, the effects that are allowed and the point at which you stop. A private plugin is not a sandbox: the write lands in the real system behind the app. If the approval you need has not come through, that gate is C; if the safe target and stop conditions do not exist yet, it is B. Don’t fire a destructive action just to see whether a confirmation prompt appears.

Custom actions are the hard case. They “do not transfer through the migration workflow,” and features that depend on them “will not work in the plugin until you set up a replacement.” That includes actions that only read data: they don’t become apps. The FAQ’s options are an available app that supports the tasks, or rebuilding the connection, which “may require a custom MCP server and technical setup.” It adds: “A rebuilt integration should not be assumed to provide every capability of the original action.”

Rebuilding is a separate project with its own permissions. Built-in migration doesn’t need the Upload plugins permissions, but rebuilding an integration “may require additional permissions.” Custom apps can be added by owners, admins and members granted custom-app or developer-mode permissions, subject to plan and workspace policy, and OpenAI says they “are not verified by OpenAI and are intended for developer use only.” If you rebuild as an imported plugin that declares MCP servers, check for a Desktop only label: such a plugin “cannot run in ChatGPT on the web.” Whether that label can apply to a migrated plugin isn’t stated.

05 People and permissions

Migrating, sharing and using the plugin are separate permissions, often held by different people.

Two points are easy to miss. Permission to use someone else’s GPT “does not give you permission to migrate it.” And migration doesn’t carry over sharing settings or give existing users access: the plugin starts private, and OpenAI’s advice is to test and review it before sharing, then ask an intended user to install and test it.

If the GPT’s creator has left, the GPT isn’t orphaned. When an owner is removed through the normal workspace flow, ownership transfers to a workspace owner and the GPT is marked unassigned; workspace owners can reassign it at any time. (“Workspace owner” is a ChatGPT role, not the business owner of the GPT.) Because a workspace admin can also migrate, a departed creator doesn’t block migration. It does leave open who maintains the plugin, which is a B until someone is named.

People outside the workspace need their own branch of the check. Public GPTs created in affected Enterprise workspaces are included in the transition even when used from a personal account or another workspace, and access to such a GPT “does not guarantee access to its replacement plugin.” The sharing options OpenAI lists for plugins are invited people or groups, anyone in the workspace with the link, or the workspace directory, and publishing to the directory “does not publish the plugin to the public directory.” None is described as reaching people outside the workspace, and the pages FSR read don’t describe another route for an Enterprise plugin. Treat that branch as C and ask your admin or account team, or as B if you’ll give those users an alternative. Settle the internal branch on its own.

06 Billing: what “no incremental cost” covers

The FAQ’s cost answer is one sentence with three conditions: “In Enterprise workspaces, using a migrated GPT without apps in Instant Chat has no incremental cost.” It then says that creating and editing plugins through a conversation use Work or Codex, and that actions in Work consume credits. It doesn’t say what the cost is measured against, or how the sentence applies under each billing model.

Your agreement decides the billing model. OpenAI says token-based billing applies to Enterprise workspaces with an eligible agreement, and adds: “Some ChatGPT Enterprise workspaces continue to use credit-based billing.” Your account team can confirm which one yours uses.

The same rate card qualifies “unlimited”: “Unlimited or virtually unlimited access is subject to abuse guardrails and any model-specific usage allowances.” And because Instant can switch to Medium, a conversation that starts in Instant Chat can still use credits on a credit-based agreement, and the FAQ doesn’t say whether “no incremental cost” covers those switched messages.

The model matters for cost as well as quality. The GPT’s selected model doesn’t carry over, and in Enterprise workspaces “Enterprise defaults apply.” OpenAI’s release notes for August 13, 2026 say owners and admins set the starting Chat model and reasoning level in Workspace settings > Models, as an admin default or the user’s last choice. Since the FAQ’s cost condition is Instant Chat, ask which one your people will land on.

On a token-based agreement, OpenAI’s rate card applies a rate to each token category and adds the results, at the rates in your agreement:

• input tokens ÷ 1,000,000 × the input rate
• cached input tokens ÷ 1,000,000 × the cached-input rate
• output tokens ÷ 1,000,000 × the output rate

The categories are separate, so count each token once. If a usage report gives you a single input figure that already includes cached tokens, splitting it wrongly is the easiest way to end up with a number that is too high. Web search, voice, image generation, long context, fast mode and regional processing can add separate charges or multipliers, and a rate that does not change with reasoning effort still leaves the number of tokens to vary. If your workspace moved from credits, OpenAI notes that “a previous credit total is not a direct estimate of the same activity’s dollar cost.” This briefing doesn’t estimate a total: that needs your rates, your usage and your integration design.

Three questions settle G5’s billing half. Put them to your account team:

1. Is our workspace on credit-based or token-based billing?
2. Does “no incremental cost” apply to our agreement, and does it cover Instant messages that switch to Medium and the Work usage that editing the plugin takes?
3. If we use data residency, does the regional processing multiplier apply to us (section 07)?

07 Data residency and sharing limits

If your workspace uses data residency, OpenAI’s residency page lists the features whose customer content is stored in the selected region: ChatGPT Memory, Code Interpreter and Data Analysis artifacts, conversations, custom GPTs with their associated prompts and outputs, files, and image generation inputs and outputs. It adds: “If OpenAI releases any new features or modalities, they will not be data residency eligible unless listed above.”

Plugins and skills are not on that list, and a search of the page on 20 September 2026, repeated on 21 September, found neither word. Whether a migrated plugin’s skill and reference files are covered, for example as “files,” isn’t stated. For a workspace with a residency requirement, that makes G5 a C: get the account team’s answer before testing with data the requirement covers.

Two settings sit behind the word residency, and they answer different questions: data residency covers where in-scope customer content is stored at rest, while inference residency, which builds on it, covers where the model runs on that content in supported regions. Neither keeps every processing step in the region. Information sent to an external app, MCP server or other third-party service “is handled under that provider’s own storage, processing, privacy, and data-residency terms.” Features that aren’t generally available, and third-party GPTs, aren’t eligible for residency at all, so ask whether plugins count as generally available in your workspace.

Two current OpenAI pages describe residency pricing differently. The residency page says data residency “is included at no additional cost for ChatGPT Enterprise and Education plans.” The token-based rate card lists “Regional processing (data residency)” at “1.1x Standard rate.” They may cover different things, a plan feature on one side and usage charges on the other, but neither page reconciles them. If both residency and token billing apply to you, ask how they combine.

Residency also limits sharing today. In workspaces with residency enabled, GPT sharing can’t be set to “Anyone,” and the page adds: “Only users within the same data residency region can access a shared GPT.” The page says this about GPTs and doesn’t say whether the same applies to plugins.

Workspace owners and admins can see the configured region on the web under Workspace settings > General, in the Data Residency Region field. This section reports what OpenAI’s pages state. It isn’t a compliance assessment.

08 Evaluate before you move users

OpenAI’s FAQ is direct about the risk: “A migrated plugin may respond differently. Compare familiar prompts and at least one harder case.” Built from its guidance, the work falls into three stages, and the line between the second and the third is the one people cross too early.

**1. Prepare, before you migrate.** Publish the edits you need; only the latest published version moves, and the GPT editor’s version history is the place to confirm what that version contains. Save a handful of prompts the GPT handles today, including at least one hard case, and write down what a good answer looks like for each. Note the conversation starters, the knowledge files and the one integration mechanism this GPT uses.

**2. Evaluate privately.** Migrate, leave the plugin private, and review the migrated instructions and reference files. Run the saved prompts against the FAQ’s four checks: the plugin selects the right skill and follows your instructions, uses the expected reference material, produces complete answers or files in the required format, and has the tools and integrations the workflow needs. Invoke it the ways your users will, since OpenAI lists an @ mention, or + and then More, where supported, and says ChatGPT may also pick a relevant installed skill automatically, but “the plugin will not necessarily run on every request.” Check which model the conversations start on (section 06) and judge the answers on that model. Test apps and rebuilt integrations separately, under the controls in section 04.

**3. Pilot with a few users, once sharing is authorized.** Sharing ends the private stage. Give a small group access under the workspace policy, ask an intended user to install it, and let them test their own permissions, which are not yours. Only then decide whether the rest of the users switch.

A pass means the cases you tried behaved well enough for the roles that tried them. It doesn’t show that the plugin matches the GPT on everything else, so draw test cases from the work that would hurt most if it failed, and include a case where access should be refused. The original stays usable, read-only, until retirement, so people can keep working while the evaluation runs; that is a fallback of last resort with an end date, not a tested rollback.

Moving everyone is its own decision. Make it after the pilot passes and the install policy is in place, and leave enough time before the retirement date in your notice to fix what the first users find.

09 How the check sorts three example GPTs

These are illustrative cases, not real deployments, and FSR hasn’t tested them.

The second example has no connected app on purpose: OpenAI’s configuration page says a GPT uses either apps or actions, not both, so a GPT that files tickets through a custom action reads its material from knowledge files rather than from a connected drive. Narrowing that evaluation isn’t a partial migration either. The whole GPT migrates, the ticket step won’t work until its replacement exists, and the pages FSR read describe no way to migrate part of a GPT.

10 FAQ

11 Methodology and limits

Tier C. For this briefing FSR used no ChatGPT Enterprise workspace, migrated no GPT, installed no plugin, connected no app and bought nothing. Nothing here measures whether migration works, how a migrated plugin behaves, or what it costs a particular workspace.

The sources are OpenAI Help Center articles and OpenAI’s Enterprise and Edu release notes, read as rendered pages, logged out and in English, on 20 September 2026 (Asia/Tokyo), reread the same evening, and checked again on 21 September, when every sentence quoted here was still on the page it is attributed to. FSR’s editor reviewed four of the cited pages on 20 September (the migration FAQ, the release notes, the token-based rate card and the residency page) and reported the same dates, cost wording and residency list; the time of that review was not recorded, and this briefing does not treat it as a separate test.

AI research assistants, ChatGPT (including a Consensus search run through it) and Kimi, helped retrieve sources, and a separate AI audit of the draft proposed corrections. Their reports were treated as leads, not findings: each quoted sentence was matched against the page it cites, the summaries around those quotes were read against the same sections, and the claims that the pages did not bear out were dropped, including one from this briefing’s own earlier draft about a GPT using an app and a custom action at once.

The French edition of the migration FAQ, marked as machine-translated, differs from the English on dates, conversation starters and planned link redirects. This briefing relies on the English edition except where the French is named.

Where this briefing says a page doesn’t mention something, that is limited to the pages cited, and was checked by reading the whole page and searching it for the relevant words. Not checked: workspace notices and admin consoles, which aren’t public; other language editions; and OpenAI’s article on administrator-managed sync. Third-party guides aren’t relied on.

These pages change often; the migration FAQ was updated within hours of the last check here. Relative update labels don’t say which sentence changed, so treat the dates, billing terms and access conditions as things to recheck before you act, along with your workspace notice. This is not legal, compliance or procurement advice.

12 Verdict

Most of the check is within your control: what moves, who can migrate and share, and what happens to custom actions are documented clearly enough to plan from. If a custom action your team needs has no replacement, fix that first and evaluate the rest around it. If migration hasn’t appeared, or your notice hasn’t confirmed the timeline, wait on that one item with a recheck date and keep preparing. If the job won’t be needed after retirement, retire the use instead of migrating it.

The limits are real. OpenAI’s own pages give two migration timelines (section 01). Undoing a migration, the limits on reference files, and whether a plugin’s parts fall under data residency are not described in the pages read, the English FAQ says nothing about conversation starters (sections 02 and 07), and cost depends on your agreement (section 06). FSR hasn’t run a migration, so nothing here says how well a migrated plugin performs. Plan from your workspace notice and your own tests, not from any single help-page date.

13 Corrections and contact

FSR BRIEFING PACKET - Is Your Custom GPT Ready to Migrate? A One-GPT Readiness Check for ChatGPT Enterprise
Publisher: Future Stack Reviews. Source: https://future-stack-reviews.com/custom-gpt-migration-readiness-check/
Evidence depth: Tier C. Public documents and pages only. No purchase, no account, no hands-on testing.
Read: 20 · 21 SEP 2026 Asia/Tokyo
Section numbers below refer to that article.
Disclosure: this briefing carries no paid link. FSR used no ChatGPT Enterprise workspace, migrated no GPT and ran no test.

WHAT IT IS
OpenAI plans to retire custom GPTs on Dec 11, 2026 and offers a migration path to plugins. In a ChatGPT Enterprise workspace the migration turns a GPT's instructions into a plugin skill, copies its knowledge files into reference files, and adds its connected apps as apps. Custom actions, conversations, sharing settings and the selected model do not carry over. The migrated plugin starts private; the original becomes read-only after migration and its creator cannot delete it. This is a six-gate check for one GPT, not a workspace-wide retirement program.

WHAT THE OFFICIAL PAGES SAY (FSR found them in disagreement)
- Migration available: Sep 22, 2026 as a target on the English migration FAQ | Sep 17, 2026 in the Sep 11 release notes | a notice on the GPT access page also targets Sep 17 and gives no creation cutoff. Section 01.
- New GPT creation ends: Oct 26, 2026 planned on the FAQ | Sep 25, 2026 in the release notes. Both are in circulation, so publish the edits you need migrated well before the earlier one.
- Retirement: Dec 11, 2026 on both pages. Every page hedges that dates may change, and a target date does not guarantee that every account can migrate that day.
- Which timeline applies to you: the notice for your account or workspace, and with several workspaces, the one where the GPT was created. Not the table above.
- Integrations: OpenAI's GPT configuration page says "A GPT can use either apps or actions, but not both at the same time." Connected apps are added as apps; custom actions do not carry over. Sections 03 and 04.
- Cost: "In Enterprise workspaces, using a migrated GPT without apps in Instant Chat has no incremental cost." The page does not say what that is measured against. Some Enterprise workspaces are on credit-based billing and others on token-based; your agreement decides. Section 06.
- Undo: the pages FSR read describe none.

UNRESOLVED - put these to your workspace admin, account team and data governance
1. The migration and creation-cutoff dates that apply to your workspace, in writing.  [why: OpenAI's two main pages give two different timelines, and both hedge.]
2. Whether your workspace is on credit-based or token-based billing, and whether "no incremental cost" still holds when a conversation switches out of Instant.  [why: the cost sentence carries three conditions and does not say what it is measured against.]
3. Whether a migration can be undone or repeated, and where your rebuild record lives.  [why: no undo is described, the original turns read-only, and its creator cannot delete it.]
4. Whether the plugin's reference files accept the same file types, counts and sizes as the GPT's knowledge files.  [why: the GPT side documents 20 files of up to 512 MB each; the pages do not say whether that carries over.]
5. Whether data residency covers every part of a plugin, and whether anyone who needs it sits outside your region.  [why: the pages FSR read do not describe the scope for plugin parts, and residency limits who a plugin can be shared with.]
(1, 2 and 5 need someone other than you. Start those first.)

FSR VERDICT
Start a private evaluation once this GPT's job, published version, integrations, people, and any billing or data rules are settled, and you accept that migrating turns the original read-only. Moving your users is a separate, later decision.
If a custom action your team needs has no replacement, fix that first and evaluate the rest around it. If migration has not appeared, or your notice has not confirmed the timeline, wait on that one item with a recheck date and keep preparing. If the job will not be needed after retirement, retire the use instead of migrating it.

LIMITS OF THIS PACKET
These pages change often; the migration FAQ was updated within hours of the last check here. Verify at help.openai.com and against the notice in your own workspace before you act. FSR used no ChatGPT Enterprise workspace, migrated no GPT, installed no plugin, connected no app and bought nothing, so nothing here measures whether migration works or how a migrated plugin behaves. This is not legal, compliance or procurement advice. Where this packet and the live pages disagree, the live pages win.

Use this briefing to tell me whether my custom GPT is ready to migrate. Ask me about the GPT's job, its integrations, who uses it, and my workspace's billing and data rules before you answer. Do not treat the dates above as settled: this briefing records where OpenAI's own pages disagree.

Read 20 September 2026, rechecked 21 September (Asia/Tokyo). Tier C: no purchase, no account, no hands-on testing. No paid links in this briefing. Not legal, compliance or procurement advice. OpenAI’s pages change often; check them and your workspace notice before acting.