On this page · 11 sections
- The dates that apply to you
- What moves to a plugin, and what does not
- How to migrate one GPT
- Rebuilding custom actions
- Public and customer-facing GPTs
- Admin steps for Business and Enterprise workspaces
- Worked example: a 150-seat Business workspace
- UK and US compliance notes
- Checklist by date
- How eCorpIT can help
- References
Summary. OpenAI's Help Center says custom GPTs are scheduled to retire on 11 December 2026 across all ChatGPT plans, 20 days from today. FedRAMP workspaces and Enterprise workspaces with a qualifying, approved deferral retire on 28 February 2027, a later date that OpenAI moved recently: an archived copy of the same FAQ from 15 November still showed 11 February 2027. At retirement, GPTs and their GPT pages become inaccessible, while existing conversations stay available. The built-in Migrate to plugin option works from a GPT's latest published version: instructions become a skill, knowledge files become plugin reference files and connected apps become apps in the plugin. Custom actions, the selected model and public sharing settings do not transfer, the new plugin starts private, and the original GPT turns read-only. Custom actions are the hard part. They need an available app or a rebuilt custom MCP server, and OpenAI lists full custom MCP support only for Business and Enterprise/Edu, with Pro limited to read and fetch. Enterprise admins can bulk-migrate through the ChatGPT Admin plugin, but bulk migration does not copy sharing settings. Our verdict: inventory this week, migrate simple GPTs before December, and start custom action rebuilds now.
About this article. This is for ChatGPT workspace owners, admins and GPT creators in the US and UK who must move custom GPTs before the retirement date. We checked dates, permissions, plan limits and prices against OpenAI's help and developer documentation, archived copies of its FAQ, the MCP Python SDK, the ICO and the European Central Bank on 21 November 2026. OpenAI tells customers to follow the notice for their own workspace, which governs if it differs from this page; this is not legal advice, and no vendor paid for inclusion. Researched with AI assistance and checked against the primary sources listed in the references.
The dates that apply to you
OpenAI's FAQ gives a standard date, a later date and Enterprise milestones, most already passed. The schedule has moved more than once. A September copy of the FAQ gave 17 September as the Enterprise migration target and 25 September as the planned end of GPT creation. Copies from late October and mid-November gave 11 February 2027 as the later date. Plan from the current page and your own workspace notice, not from an older admin email.
| Date | What happens | Who it applies to |
|---|---|---|
| 11 September 2026 | Admin notice sent | Affected Enterprise workspaces |
| 22 September 2026 | Target for the migration experience and user banner (a target, not a guarantee) | Affected Enterprise workspaces |
| 1 October 2026 | Target for the migration banner | Enterprise workspaces that opted into delayed banners |
| 26 October 2026 | New GPT creation ended; edits to unmigrated GPTs still allowed | Affected Enterprise workspaces, except FedRAMP |
| 11 December 2026 | Standard retirement: custom GPTs and their pages become inaccessible | All ChatGPT plans unless your notice says otherwise |
| 28 February 2027 | Later retirement | FedRAMP workspaces and approved deferrals only |
Three caveats from the FAQ are easy to miss. The standard admin emails carry the standard dates and are not tailored to workspaces with approved delays. A deferral on its own does not keep the right to create new GPTs. And the date follows the workspace where the GPT was created, not the plan of the person using it: a public GPT built in an affected Enterprise workspace follows that workspace's timeline even when someone opens it from a personal account.
US public-sector teams: unless your workspace is FedRAMP or holds a written deferral from your account team, plan for 11 December. The FAQ gives no separate date for UK organisations. If you also used Agent Builder, its shutdown on 30 November comes first; see our note on the Agent Builder, Evals and prompts shutdown.
What moves to a plugin, and what does not
A ChatGPT plugin can package skills (reusable instructions), connected apps, app templates and extensions. Developers build them from skills, MCP servers and optional interfaces. The table summarises what migration keeps, from OpenAI's FAQ and admin guide.
| GPT component | After Migrate to plugin | Action needed |
|---|---|---|
| Instructions (latest published version) | Become a skill in the plugin | Publish final edits first; drafts and unpublished edits do not transfer |
| Knowledge files | Copied into the plugin's reference files | Check the files are complete and still current |
| Connected apps | Added to the plugin as apps | Each user still needs app access and account authorisation |
| Custom actions | Not transferred | Rebuild with an available app or a custom MCP server |
| Selected model | Not carried over; Enterprise defaults apply in Enterprise workspaces | Test answers on the default model |
| Sharing and audience | Plugin starts private; public sharing settings dropped | Reshare; public plugins need a separate submission |
| Scripts used by tools | Standard Chat, including Instant, does not currently run scripts bundled with plugins | Test any code or file generation where it will be used |
| Conversation starters | May not copy one to one | Save the prompts you rely on |
| The original GPT | Read-only after migration; its creator cannot delete it | Make all future changes in the plugin |
| Past conversations | Remain accessible after retirement | Cannot be moved into a project |
Finish edits before migrating, because the GPT freezes. The GPT page also disappears at retirement. The September copy of the FAQ said migrated GPT links were planned to redirect to the replacement for people with access; the current FAQ says only that GPT pages become inaccessible, and the admin guide says retired GPTs leave the GPT directory. If your website, intranet, help centre or sales emails link to a GPT, replace those links.
The admin guide also says a GPT can still be migrated after retirement, although it will no longer run. That is a safety net for a forgotten GPT, not a plan for anything people use daily.
When a project is the better replacement
The FAQ suggests a project for GPTs used to keep instructions and files together for ongoing work such as writing or research. A project does not take over custom actions or sharing settings.
| Choose this | If the GPT |
|---|---|
| Migrate to plugin | Is shared with a team, uses connected apps, or is reused across many chats |
| Project | Was a personal workspace for one ongoing piece of work with a few reference files |
| Plugin plus rebuilt MCP server | Used custom actions to read or write data in your own systems |
| Plugin submitted to the public directory | Is customer-facing and was public in the GPT Store |
| Retire without replacement | Shows little or no use in the last 30 days |
In Enterprise workspaces, admins with the ChatGPT Admin plugin can request a GPT report with names, owners and message counts for the last 30 days, which is the quickest way to find GPTs nobody uses.
How to migrate one GPT
For a GPT you created, OpenAI's process is short.
- Finish and publish any essential edits. Publishing does not require public sharing.
- Save three to five familiar prompts and one harder case, with the answers the GPT gives today.
- Go to My GPTs, find the GPT under Created by me and select Migrate to plugin. If the option is missing, migration may not be available to your account yet, plugins may be disabled, or you may be in a different workspace from the one where you created the GPT.
- Review the migrated skill and reference files.
- Run your saved prompts. Check that the plugin selects the right skill, uses the expected reference material, produces complete answers or files in the required format, and has the tools the workflow needs.
- Share the plugin. In a managed workspace, Share Plugin can prefill the original GPT's users, groups or workspace audience; review the list and select Save.
- Ask one intended user to install and test it, then tell everyone else to switch.
Users reach the plugin with an @ mention, by opening + and selecting More where supported, or automatically when ChatGPT judges an installed skill relevant. Automatic selection depends on the request, so tell people to mention the plugin by name for important tasks.
Rebuilding custom actions
Custom actions take most of the migration time. OpenAI says they do not pass through migration, and features that depend on them stop working until you set up a replacement: an available app that supports the task, or a rebuilt connection that may need a custom MCP server.
Plan availability decides what is possible.
| Plan | Custom MCP server support | Who can create and publish |
|---|---|---|
| Enterprise and Edu | Full MCP, including write actions | Admins and owners, plus members granted the permission through RBAC; only admins and owners publish |
| Business | Full MCP, including write actions | Only admins and owners create and publish |
| Pro | Read and fetch only for custom MCP servers; can build apps with the Apps SDK | Individual user |
| Plus, Go and Free | Not listed for custom MCP servers | Consider a plugin with an MCP app hosted on ChatGPT Sites |
The Sites route needs care. OpenAI says a plugin with an app hosted on ChatGPT Sites can be created on any ChatGPT plan, but Pro and personal-account users cannot currently share a Site-hosted plugin with other ChatGPT users, and Sites is included with eligible plans during its public beta. That suits a personal tool; a customer-facing GPT needs a Business or Enterprise workspace or a public plugin submission.
A minimal rebuild pattern
Most GPT actions call an existing HTTPS API described by an OpenAPI schema. The cleanest rebuild is a small remote MCP server that wraps the same API, exposes one tool per operation and keeps credentials on the server. Version 2 of the official MCP Python SDK renamed the v1 FastMCP class to MCPServer. This read-only example uses only the SDK and the Python standard library, and URL-encodes the order ID so a model cannot alter the request path:
import json
import os
import urllib.parse
import urllib.request
from mcp.server import MCPServer
mcp = MCPServer("order-status")
@mcp.tool()
def get_order_status(order_id: str) -> dict:
"""Read-only: return status and ETA for one order ID."""
url = f"{os.environ['ORDERS_API']}/orders/{urllib.parse.quote(order_id, safe='')}"
req = urllib.request.Request(
url, headers={"Authorization": f"Bearer {os.environ['ORDERS_TOKEN']}"}
)
with urllib.request.urlopen(req, timeout=10) as resp: # raises on 4xx and 5xx
return json.load(resp)
Install the SDK with its command-line extra and serve the file over Streamable HTTP, as the SDK's README shows:
uv add "mcp[cli]"
uv run mcp run server.py --transport streamable-http
The README's client example then connects at http://localhost:8000/mcp. Put the server behind your normal HTTPS ingress and authentication. To add it in ChatGPT, go to Plugins, select the plus button and Add custom MCP server, enter the server's public HTTPS endpoint or choose a Secure MCP Tunnel, configure authentication, acknowledge the risk warning and select Create as a plugin. Five details from OpenAI's documentation will save rework:
- OAuth refresh tokens. If your identity provider does not issue refresh tokens, usually through the
offline_accessscope, ChatGPT may lose access when the first authorisation expires and users must sign in again.
- Frozen tool snapshots. After an admin approves a server, ChatGPT uses a frozen snapshot of its tools. Incompatible changes to tool definitions cause errors until an admin refreshes them, and admins are not notified automatically.
- Write actions. ChatGPT may ask users to confirm writes, may block especially risky actions, and shows admins risk warnings. On Enterprise and Edu, admins can switch individual actions on or off, and new actions arrive disabled.
- Private systems. ChatGPT connects only to remote MCP servers. For on-premises or private-network systems, use Secure MCP Tunnel rather than exposing the server publicly; our Secure MCP Tunnel setup guide covers the client.
- Surfaces. Custom MCP servers are connected on the web, not mobile, and deep research can use them for read and fetch actions only.
OpenAI also warns that connecting untrusted MCP servers raises exposure to prompt injection, and that a rebuilt integration may not offer everything the original action did. Budget time for a security review of every write action.
Public and customer-facing GPTs
If customers use a GPT you published in the GPT Store, migration alone will not keep them served. The plugin starts private, public sharing does not carry over, and access to the old GPT does not grant access to the replacement. Making a personal plugin public requires OpenAI's separate submission process on the OpenAI Platform. An organisation owner, or a member with Apps Management Write, uploads a plugin ZIP, completes individual or business verification to publish under a personal or company name, resolves automated findings and submits for review. Remote MCP servers must use a stable, public HTTPS endpoint.
One detail changes the order of work. OpenAI says adding an MCP server to an existing skills-only plugin is not currently supported, so if the public plugin needs a rebuilt action, include the MCP server in the first ZIP rather than submitting the skills first and adding it later.
Admin steps for Business and Enterprise workspaces
OpenAI's detailed admin guidance is written for Enterprise workspaces; Business admins should follow their in-product notice where it differs. Permission names below are OpenAI's.
- Check plugin permissions under Workspace settings > Permissions & roles. Use plugins is on by default for Enterprise; Upload plugins and Create plugins with MCPs are off by default. The built-in migration does not need the upload permissions. Share plugins and Publish plugins to workspace are separate permissions.
- Build the inventory. Retrieving the full workspace GPT inventory currently requires a workspace Owner. Record creator, audience, knowledge files, connected apps, custom actions and recent use for each GPT.
- Decide owners. In Enterprise, only a GPT's creator or a workspace admin can migrate it. For GPTs whose creators have left, assign an admin now.
- Bulk-migrate where it helps. The ChatGPT Admin plugin offers bulk migration. It runs asynchronously, so starting the job does not confirm success; review the results and test each plugin. Bulk migration does not copy sharing settings, so record each GPT's audience first.
- Set installation policy. For each plugin, choose Available or Installed and confirm required apps are enabled. Admins can change availability for up to 25 plugins at a time.
- Reshare. Migrated plugins start private. Use the audience prefill in Share Plugin where available, review it and save.
- Rebuild actions with the owning engineering team and a security review, using the MCP steps above.
- Communicate twice: when replacements are ready, and a week before 11 December with a list of GPTs that will stop.
On credits, the FAQ says using a migrated GPT without apps in Instant Chat has no incremental cost in Enterprise workspaces, while creating and editing plugins through a conversation uses Work or Codex, and actions in Work consume credits. The admin guide adds that the conversion itself consumes no credits, that a request in Instant Chat can switch to a reasoning model that does, and that Enterprise agreements billed in US dollars have different terms, including charges for Instant usage. Check your rate card before rollout.
Worked example: a 150-seat Business workspace
The counts and rates below are assumptions, not measured data; replace them with your own.
Assumptions. A US and UK software company has 150 ChatGPT Business seats on annual billing. OpenAI lists Business at $20 per user per month billed annually, or $25 billed monthly, so the seats cost $3,000 a month (£2,285.76). Our business AI pricing comparison sets that against rival plans. The inventory finds 36 custom GPTs: 24 with instructions and files only, 8 with connected apps, 2 with read-only custom actions and 2 with write actions that create support tickets. We assume an internal blended engineering and admin rate of $110 an hour (£83.81). Pound figures use the European Central Bank reference rates for 20 November 2026 (US$1.0990 and £0.83735 to the euro), crossed to £0.76 per dollar and rounded to two decimals.
| Work item | Count | Hours each (assumed) | Hours | Cost (USD) | Cost (GBP) |
|---|---|---|---|---|---|
| Instruction and file GPTs: migrate, test, reshare | 24 | 1.5 | 36 | $3,960 | £3,017.20 |
| GPTs with connected apps: migrate, test app access | 8 | 3 | 24 | $2,640 | £2,011.47 |
| Read-only custom actions: MCP rebuild and test | 2 | 16 | 32 | $3,520 | £2,681.96 |
| Write custom actions: MCP rebuild, OAuth, security review | 2 | 32 | 64 | $7,040 | £5,363.92 |
| Total | 36 | 156 | $17,160 | £13,074.55 |
The arithmetic: 24 × 1.5 = 36 hours; 8 × 3 = 24; 2 × 16 = 32; 2 × 32 = 64; total 156 hours × $110 = $17,160. The four GPTs with custom actions take 96 of 156 hours, about 62% of the effort, although they are 11% of the GPTs (4 of 36). At 156 hours, one person cannot finish in the 20 days left, so spread simple migrations across GPT creators and give engineers the rebuilds. Hosting costs for the MCP server depend on where you run it and are not included.
UK and US compliance notes
A rebuilt action can change how personal data flows: a new server, new credentials, perhaps new hosting. For UK organisations, review whether your data protection impact assessment and records of processing still describe the integration accurately. The ICO says its DPIA guidance is under review after changes made by the Data (Use and Access) Act, so check the current version when you update yours. OpenAI also notes that data sent to a connected service falls under that provider's own storage, privacy and residency terms, and that data residency does not keep every external integration in the selected region.
In regulated US sectors, put each rebuilt write action through your normal change and access review. OpenAI says Enterprise and Edu conversations, including those that use apps, are available through its Compliance API, but that log coverage depends on the app, product and workspace configuration, so confirm it per app before relying on it as evidence.
Checklist by date
- By 26 November: finish the inventory, assign an owner to every GPT, record each audience, and freeze edits on GPTs you will migrate.
- By 1 December: migrate and test all instruction-only and app-based GPTs; reshare them; update links on websites and intranets.
- By 8 December: finish custom action rebuilds, complete security review, publish MCP-based plugins and test with real users.
- 11 December: confirm what stopped and message users of GPTs retired without replacement. Other services also cut off this month; see our December developer cutoffs list.
- Before 28 February 2027: FedRAMP and approved-deferral workspaces repeat the same steps against the later date.
How eCorpIT can help
eCorpIT is a Gurugram-based technology organisation, founded in 2021, assessed at CMMI Level 5 and MSME certified, with senior-led engineering teams working across AWS, Microsoft and Google platforms. We inventory custom GPTs, sort quick migrations from rebuilds, and build and test the remote MCP servers that replace custom actions through our custom MCP server development service. Talk to us at /contact-us/.
Last updated: 21 November 2026.