Notes on coding agents and the pages they leave behind, on moving between notes apps, and on the plain mechanics of keeping what you think. Written and published from a Sumibako vault.
Custom GPTs retire on 11 December. What moves, what breaks, and how to keep a copy
Custom GPTs retire on 11 December 2026. What the move to plugins keeps, what it drops, and how to save a copy of your GPT before it goes.
Prashant Shaw
Founder, Sumibako
8 min read
On 11 September OpenAI announced that custom GPTs are retiring. The date is 11 December 2026, and after it, in OpenAI's words, "custom GPTs and their GPT pages become inaccessible." Somewhere a GPT called Recipe Wizard 3000 has seven weeks left and does not know it.
If you built a GPT, use one every day, or sell access to one, this is what those seven weeks look like: what the official move to plugins carries over, what it leaves behind, what people in OpenAI's forum have already run into, and how to keep a copy of your GPT that no retirement can reach. Everything here comes from OpenAI's help pages and forum, checked on 22 October 2026. I build a notes app, which turns up once near the end.
🏖️ What is happening, and when
The release notes for 11 September put it in one line: "We’re planning to retire custom GPTs across ChatGPT plans and provide a migration path to plugins". The retirement FAQ fills in the rest:
When
What happens
Already
Personal accounts cannot create or publish new GPTs. Existing ones still work, and can still be edited where your plan allows
Custom GPTs and their pages stop working, on every plan
11 February 2027
The same, for Enterprise workspaces with an approved deferral
The first row is from OpenAI's page on creating and editing GPTs: "New GPT creation and publishing are not available on personal ChatGPT accounts, including Free, Go, Plus, and Pro." Business, Enterprise and Edu workspaces could still make them when I checked, though the FAQ's Enterprise timeline ends that on 26 October.
Two pieces of good news from the FAQ. Your old chats survive: "Your existing conversations with custom GPTs will remain accessible after retirement. You don't need to take any action to preserve them." And if you only use somebody else's GPT, there is nothing for you to migrate: "You do not need to migrate a GPT you only use".
NOTE: These are OpenAI's planned dates, and the FAQ's own timeline says "the dates are subject to change". It also says the notice in your account is the one to follow, so check that before you plan around anything here.
🧳 What the move carries, and what it drops
When migration reaches your account, you open My GPTs and select Migrate to plugin. Here is what happens to each part of your GPT, from the FAQ:
Part of your GPT
After you migrate
Instructions
Become a skill inside the new plugin
Knowledge files
Copied into the plugin's reference files
Connected apps
Added to the plugin as apps
Custom actions
Do not transfer. You rebuild them, possibly as an MCP server
The model you chose
Does not carry over
Who you shared it with
Does not carry over. The plugin starts private
Drafts and unpublished edits
Do not transfer
The original GPT
Keeps working until retirement, but becomes read-only, and you cannot delete it
Three of those rows deserve their own sentence.
Drafts. "Migration uses the GPT's latest published version. Drafts and unpublished edits do not transfer." If you have been tinkering without pressing Update, the plugin gets the version from before the tinkering.
Actions. "GPT custom actions do not transfer through the migration workflow." If your GPT calls an API, the FAQ says rebuilding it "may require a custom MCP server and technical setup", and warns that a rebuilt integration "should not be assumed to provide every capability of the original action."
Sharing. "A migrated personal plugin starts private. Making it public requires a separate plugin submission process." People who used your GPT do not follow it to the plugin.
The FAQ does not mention conversation starters either way, so copy them out yourself. They take a minute to save and an annoying afternoon to remember.
🚧 What people have already run into
Migration has been reaching accounts since late September, and OpenAI's forum has the first reports. These are individual users, not OpenAI, so treat them as things to check rather than things that will happen.
Knowledge files refused. One user reported the migration asking them to replace knowledge files because they were "too large", files "accepted by the GPT". The same person then added the same PDF to a plugin built from scratch, and "The plugin update succeeded". If the migration refuses a file, adding it by hand afterwards is worth a try.
Paying customers left behind. The busiest thread, at over 3,100 views, is from a small creator asking: "How are small creators supposed to provide continued access to customers who have already paid for our GPT-based products?" The FAQ's answer is the line about sharing above, plus: "Access to the original GPT does not automatically give you access to a replacement plugin."
A different kind of thing. Another user put it this way: "this does not look like a migration. It looks like a change of product model." A GPT was a room you walked into. A plugin is a tool ChatGPT picks up when it thinks it needs one, and the FAQ says as much: ChatGPT uses an installed skill "when its description matches your request", and "the plugin will not necessarily run on every request." You can still call it by name with an @ mention.
That last point matters more than it sounds. If your GPT's value was that every message in it followed your instructions, test that the plugin does the same before you rely on it.
📋 Before you press Migrate
1️⃣ Press Update. OpenAI's editing page says that when editing an existing GPT you "select Update to apply your latest draft changes". Do it before migrating, since drafts do not transfer.
2️⃣ Copy everything into one file you own. The instructions word for word, the description, the conversation starters, which capabilities are on. A template is in the next section.
3️⃣ Find the original knowledge files. OpenAI's help pages describe adding a file to a GPT and removing one in the GPT builder. I could not find one that describes downloading a file back out. If the only copy of your handbook lives inside a GPT, fix that first.
4️⃣ Write down your tests. The FAQ suggests you "Compare familiar prompts and at least one harder case." Pick five prompts you really use, save the answers you get today, and you have something to compare the plugin against.
5️⃣ List your actions. For each one: what it does, which API it calls, how it signs in. That list is the brief for whoever rebuilds it, including future you.
6️⃣ Then migrate, test, and only then tell anyone to switch.
📄 Your GPT, as one file
This is the whole GPT in a form any assistant can read. Fill it in from the GPT's configuration view (My GPTs, then Edit GPT):
One file is the right shape because it is the shape every replacement wants. A plugin's skill is instructions. A Claude project is instructions plus files. Anthropic publishes its skill format as an open standard, and Google, it turns out, is retiring its own version of custom GPTs: "Starting in November 2026, Gems will transition to skills in Gemini Apps for personal Google Accounts" (Google's page). The custom chatbot is becoming a skill everywhere at once, and a skill is a document.
🧭 Where it can live after 11 December
Option
What it holds
Worth knowing
A ChatGPT plugin
Your instructions as a skill, your files as reference files
The official route. Creating plugins is a Go, Plus and Pro feature on the pricing page, not a Free one
A Claude project
Project instructions plus a knowledge base, used by every chat in the project
The closest thing to a GPT's room. Free accounts get up to five (Anthropic)
A Claude skill
Instructions and files Claude loads when a task needs them
Anthropic calls skills "folders of instructions, scripts, and resources"
A Gemini Gem
Instructions plus knowledge files
Turning into a skill from November, by Google's own timeline
A page you keep
The file from the previous section
Nothing runs it. You paste it, or connect an assistant that can read it
If you are moving to Claude anyway, a project is the easiest landing: paste the instructions into the project's instructions, upload the knowledge files, and every chat in that project starts where your GPT did. A project even keeps a memory of its own. I wrote a separate guide to moving from ChatGPT to Claude for everything else that comes with it.
📍 Keep the master copy somewhere that cannot retire
The lesson of 11 December is that a GPT was always a document living in somebody else's house. Plugins, projects, skills and Gems will all change again. The file from this post will not, as long as it lives somewhere you control: a folder, a repository, a note.
That last one is where I come in, so weigh this paragraph accordingly. Sumibako is a notes app that AI assistants can read and write pages in. Keep your GPT's file there and Claude can read it on every plan, Free included, through a custom connector, so "use my GPT page as your instructions" works in any new chat. On ChatGPT, OpenAI's page on custom servers allows reading on Pro and reading and writing on Business, Enterprise and Edu; on other plans you paste it. Sumibako does not run anything and is not a GPT replacement. It just keeps the instructions where no retirement can reach them.
Custom GPTs are retiring to the beach on 11 December. Pack their bags before they go.
Eight eMudhra keys were integers under 256, one of them 4. Why browsers blocked them in hours, why your Docker image did not, and what to check by 27 October.
10 min read
markdown
# GPT: <name>Description: <the one line people see before opening it>Recommended model: <if you set one>Capabilities: <web search, image generation, canvas, code interpreter, apps>## Instructions<the full Instructions field, pasted word for word>## Conversation starters- <starter one>- <starter two>## Knowledge files- <file name>: <where the original lives>## Actions- <name>: <what it does>, <the API it calls>, <how it signs in>## Tests- Prompt: <a prompt you really use> Good answer looks like: <what you expect back>
