Source: ai-research/openai-custom-gpt-retirement-faq-2026-09.md — Project: OpenAI Help Center, “Custom GPT retirement and migration FAQ” · URL: https://help.openai.com/en/articles/20001519-custom-gpt-retirement-and-migration-faq · fetched 2026-09-29 from the ur-in locale copy, because the en URL refused the crawler. Commentary from raw/Ep._243_-_GPT-6_Sol_Luna_Opus_5.5_New_Microsoft_Copilot_Huang_v._AI_Doomers_Intro_to_AI_Score.md (The Artificial Intelligence Show Ep 243, Mike Kaput and Paul Roetzer, late September 2026, https://www.youtube.com/watch?v=vJl6AaQ6B0A) and raw/Ep._236_-_AI_Answers_-_No_Time_for_AI_AI_Budgets_Vendor_Terms_Data_Risk_AI_Disclosure.md (Ep 236, Roetzer and Cathy McPhillips, https://www.youtube.com/watch?v=NLFQkHiaCiQ).
OpenAI is retiring custom GPTs. For affected Enterprise workspaces they are scheduled to stop running on 2026-12-11, and OpenAI says the transition affects all ChatGPT plans. The replacement is the plugin: a package of skills (reusable instructions) and apps (connections to services). A built-in migration flow turns a GPT’s instructions into a skill, but custom actions, the chosen model and sharing settings do not carry over, and public distribution now goes through an OpenAI review. Anyone who built GPTs for clients, courses or lead magnets has a hard deadline and some rebuilding to do.
Key Takeaways
-
Scope. “The transition affects all ChatGPT plans.” The published dates and admin guidance describe the Enterprise transition, and “other plans may follow the same transition timeline.” Public GPTs created in affected Enterprise workspaces are included even when used from a personal account or another plan: “What matters is where the GPT was created.”
-
Dates for affected Enterprise workspaces (OpenAI: “the dates are subject to change”):
Date Milestone 2026-09-11 Admin notice 2026-09-17 Target for the migration experience and user banner (“not a guarantee”) 2026-09-25 (planned) Creation of new custom GPTs ends; publish any drafts you need to migrate before this 2026-12-11 Scheduled retirement: custom GPTs stop running and leave the GPT directory Kaput, reading what he called the current FAQ on air in late September, gave the creation cutoff as October 26: “For affected workspaces, OpenAI now plans to stop new GPT creation on October 26th.”^[ambiguous] See Open Questions.
-
What transfers. The GPT’s instructions become a skill within the new plugin; connected apps become apps in the plugin; reference files, templates and examples come across but need review (Kaput: knowledge files become the plugin’s reference files).
-
What does not transfer:
- the selected model (Enterprise defaults apply)
- custom actions (“rebuild required integrations using a supported connector or custom MCP server”)
- conversation starters and previous chats (“may not copy”)
- drafts and unpublished changes: only the latest published version migrates (Kaput)
-
Who can migrate: the GPT’s creator or a workspace admin, with plugins enabled, on a published GPT. “Permission to use someone else’s GPT does not give you permission to migrate it.”
-
After migration the original stays usable until retirement but becomes read-only, and its creator cannot delete it. Migrated GPT links are planned to redirect to the replacement, but “a redirect does not grant access to the plugin.”
-
Sharing starts from zero. “The replacement plugin starts private.” Sharing with people or groups and publishing to the workspace directory are separate permissions (Share plugins; Publish plugins to workspace). Kaput adds that you “cannot share these publicly”: public listing means submitting the plugin through a portal for OpenAI review, with no stated timeline.
-
Behaviour can change. “A migrated plugin may respond differently.” OpenAI’s own test list: compare familiar prompts plus at least one harder case, and check that the plugin selects the right skill and follows instructions, uses the expected reference material, produces complete answers or files in the required format, and has the tools and integrations the workflow needs.
-
How users invoke it. Select the plugin or skill with an @ mention, or open + and choose More. ChatGPT “may also use a relevant installed skill automatically” when its description matches the request, but “the plugin will not necessarily run on every request.”
Commentary — The Artificial Intelligence Show
Opinion from the Ep 243 hosts, labelled as such:
- Roetzer: the easy on-ramp is gone. GPTs launched at DevDay in November 2023. OpenAI “never seemed to commit to the original vision… building this entire marketplace.” Yet GPTs were “the fastest way to teach someone how to get value out of AI.” His own JobsGPT has had 44,000 conversations and 800+ ratings. “They took GPTs, which were easy to understand, build and share, and blew it up.”
- Roetzer: plugins and agents don’t mix yet. As of recording, you could connect an app to a ChatGPT agent but not a plugin, so a GPT rebuilt as a plugin could not be called from an agent (“maybe they fix this”).
- Roetzer’s three options:
- rebuild the GPT as a standalone app on the API
- create and share a plugin, if sharing works for your case
- just give people the prompt, which his company leans toward because course attendees use Copilot, Claude or Gemini too
- Kaput: start with skills. “If you’re not using skills yet, which are just markdown files that point at folders and examples and files, start there and then turn that into a plugin. Don’t even worry about the app piece of it yet.” A GPT “took 2 seconds”; the plugin route means creating a skill, packaging files, wiring apps and bundling it all, which is “much more onerous.”
- Kaput on client GPTs: for people who built public GPTs for clients, “the solution is you’re screwed right now.”
The Platform-Risk Lesson (Ep 236, recorded early September)
Cathy McPhillips asked whether businesses built on custom GPTs repeat the “rented land” mistake of building audiences on social platforms. The episode says MAICON (October 13) is “41 days” away, which dates it to about 2026-09-02, before OpenAI’s September 11 admin notice.^[inferred] Roetzer’s answer reads as a migration plan:
- System instructions from a GPT “can very easily be copied and pasted into Claude.”
- Document every app and agent: its system instructions and the knowledge base it uses, so you can audit and rebuild it. His company now does this as part of governance.
- Intelligence redundancy: at the extreme, keep a mirrored setup on a second platform “so that if something goes down you just flip a switch.”
Try It
- Inventory now. List every GPT you rely on or publish: owner, audience, instructions, knowledge files, custom actions, and a few familiar prompts with good outputs to test against. OpenAI recommends exactly this preparation.
- Publish any drafts you intend to keep before the creation cutoff (Sep 25 or Oct 26; confirm in your workspace). Only published GPTs migrate.
- Triage custom actions first. They are the only piece that needs real rebuilding, as a supported connector or a custom MCP server. Allow time for security review.
- Export the instructions as a portable skill file you control. The same text can seed a Claude skill or project if you want vendor redundancy (Roetzer’s copy-paste point).
- Migrate, test, then share. Run the familiar prompts plus one harder case through the plugin, then share it with the audience explicitly. Nothing is shared by default.
- For public or client-facing GPTs, plan a replacement. Lead magnets, course tools and client assistants need one: a published prompt, a hosted app, or a submitted plugin whose approval timing is unknown.
Open Questions
- Which creation cutoff is current? The fetched first-party FAQ says 2026-09-25 (planned). Kaput, reading what he called the current FAQ in late September, said October 26. The crawled copy carries only a relative “last updated” stamp, so it may predate a change. Check the live English FAQ or your in-product notice.
- Dates for non-Enterprise plans (Plus, Pro, Business, personal accounts) are not published. OpenAI says they “may follow the same transition timeline” and to watch in-product notices.
- Public plugin review has no published criteria or turnaround time, which matters for anyone distributing GPTs to an audience.
- Can a plugin be attached to a ChatGPT agent? Roetzer said no as of recording; unverified since.
Related
- Agent Skills Overview (Anthropic) — the Claude-side skill format, if you rebuild portably.
- Claude Code Plugins and Marketplaces — Anthropic’s equivalent plugin packaging.
- Distributing Team Skills as a Plugin Marketplace — sharing skills across a team without a vendor directory.
- Lead Magnet Creation with Claude Code — recommends a custom GPT as a lead-magnet format; affected by this retirement.
- The AI Tools Assessment (Corey Gannon) — cites a custom-GPT knowledge system as an example; affected.
- OpenAI Enterprise Signals — OpenAI’s own data on who uses plugins and skills.
- AI Transformation Case Studies (AI Show) — the same hosts’ leader Q&A on vendor risk and redundancy.