Custom GPTs shut down on December 11: How to save your own bots

What's going on

OpenAI is retiring a feature that has been part of ChatGPT since late 2023: Custom GPTs. These are custom-built versions of ChatGPT with their own instructions, their own knowledge files and sometimes connections to external services. According to the trade publications Campus Technology, Virtualization Review and THE Journal, OpenAI announced the move on September 11, 2026. The standard shutdown is scheduled for December 11, 2026. OpenAI's official help page notes that the timeline and available migration features may vary by account and workspace.

GPTs are being replaced by plugins. These extensions work in both ChatGPT and Codex, OpenAI's coding agent. According to the reports, Enterprise workspaces affected by the transition already lost the ability to create new GPTs on October 26. Enterprise customers with an approved extension and FedRAMP environments (FedRAMP is a US security standard for government clouds) can keep using their GPTs until February 11, 2027.

So if you've built GPTs or share them with others, now is the time to act. On November 7, 2026, a writer on the Japanese platform note.com walked through how to save your own GPT. She tested three routes: as a ChatGPT plugin, as a skill in Claude, and as a bot on the Poe platform.

What changes under the hood: plugins, skills and MCP

Until now, every Custom GPT was its own self-contained version of ChatGPT. Plugins are built from reusable components instead. According to the reports, a plugin can contain two things. One is skills: bundles of instructions and supporting material for recurring workflows. The other is an MCP server. The Model Context Protocol is an open standard that lets an AI access tools and external systems, such as company databases. A plugin can also contain both. ChatGPT and Codex share a common plugin directory, but some features are only available in one of the two environments.

OpenAI's developer documentation describes plugins as packages you can "discover, install, share and publish." OpenAI has already reworked the interface: GPTs are less prominently placed, and if you try to create a new GPT, you're offered to create a plugin instead.

What gets lost in the move

Under "My GPTs" there's a button to convert an existing GPT. Your instructions become a skill, knowledge files end up as reference files in the plugin, and connected apps can be carried over. But plenty doesn't move automatically:

• Past chats with the GPT are not transferred.
• The selected model is not carried over.
• Custom Actions, i.e. your own API integrations, may need to be rebuilt, either via an available app or your own MCP server.
• Sharing settings are lost. Migrated plugins start out private, existing users of a GPT don't automatically get access, and old sharing settings no longer apply.

OpenAI itself also warns:

"A migrated plugin may respond differently."

The company recommends re-testing your usual prompts, how reference material is handled, required output formats and integrations. According to Converge 360's reporting, some GPT builders are pushing back. They argue that plugins can't fully replicate the self-contained, predictable environment of a Custom GPT. Other criticism concerns sharing, the effort of migrating and the changed way instructions and material are processed.

Route 1: Stay in ChatGPT as a plugin

For the hands-on test, the author used her simple GPT "What Day Is It Today." Given an observance day and an image, it writes a short Instagram Reels caption. The rules: no more than 200 characters, a bright, friendly tone, two mandatory hashtags (#TodayInHistory, #Midjourney) and no more than two additional hashtags. At the end, the GPT asks "Does this look okay?" and after an "OK" it asks for the next image. The GPT had no knowledge files or external integrations. Complex GPTs were explicitly not tested.

Her most important advice: Before migrating, back up the name, description and instructions, take screenshots of the enabled features and keep a standard test run with a fixed image and a fixed prompt. She used a kitten in a sweater with a mug of cocoa for "Cocoa Day" on November 7.

The conversion itself was quick. A confirmation dialog pointed out that the plugin would be created as private and that the original GPT could no longer be edited afterwards. When tested with "Try in chat," the caption was worded differently but still fit Cocoa Day. The hashtag rules were followed, and the flow of asking for confirmation and then the next image worked too. Her takeaway: Don't check whether you get the same text. Check whether you can carry on with your usual work.

Route 2: As a skill in Claude

From the migrated plugin's menu, you can download a ZIP archive. The author didn't upload this ZIP directly to Claude. Instead, she handed it over in a regular Claude chat (not Claude Code) with precise instructions: review the contents, don't change the instructions and rules, only adjust the file structure, name and description, explain unsupported features instead of silently dropping them, and finally deliver an uploadable ZIP along with a list of changes. If you struggle to phrase this, the author says you can have ChatGPT write the request for you first.

Claude delivered the ZIP and an explanation of the changes. You upload it in the settings under "Skills" → "Upload Skills." According to Claude's official help docs, code execution and file creation need to be enabled. The author explicitly selected the skill in the input field. After that, the workflow ran as usual.

One difference stood out, though: the caption came back as plain running text and was awkward to copy. Using "Edit with Claude," she added an instruction to put the post in a code block. After that, a copy button appeared, and the follow-up question sat outside the code block. Her lesson: Don't just check the content. Check how usable the output is in day-to-day work.

Route 3: Share with others via Poe

If you've shared GPTs with other people, you need a new home for them. The author tested Poe, a platform where you can use models from multiple providers, including GPT, Claude and Gemini. There she created a "Prompt Bot": pick a name, description and base model, paste in the old GPT instructions and set access to "Only you." Heads-up: The save button is labeled "Publish," but the bot was still private afterwards.

Cost is the sticking point. Poe bills in its own points, and a ChatGPT or Claude subscription doesn't help you there. Free users get a daily points allowance. In the test, the points weren't enough for GPT-5.4, and GPT-6.1-Terra would have required 701 points. Only with GPT-4o-mini did the bot run, and the balance dropped from 300 to 131 points, a cost of 169 points. The author stresses that this figure applies only to this single run. Consumption depends on input and output length, the model and features such as web search.

On the content side, GPT-4o-mini made up a small detail: "The cat is enjoying the cocoa, too" wasn't actually shown in the image. So check captions before you post them.

Poe's advantage: You can edit the bot centrally instead of sending every user new instruction files. You can share it with "Everyone," "Only people with the access link" or "Only you." According to Poe's FAQ, you get the link by clicking the bot's name and then "Share bot." The author didn't verify whether third parties can actually use the bot via the link. Poe also offers a "Knowledge base" for TXT, PDF, DOCX and HTML files, but you have to re-upload the files yourself, and that wasn't tested either. ChatGPT itself also lets you share plugins. The conditions are laid out in OpenAI's migration guide.

Our take

The shutdown affects a lot of people who have built small helpers over the past three years, from hobby bots to company tools. For simple GPTs that consist only of instructions, the hands-on test shows the move is straightforward, even across providers. The fact that ChatGPT skills can be used in Claude with a bit of rework shows how similar the big providers' extension concepts have become.

The tricky part is anything beyond plain instructions: Custom Actions, knowledge files and especially publicly shared GPTs. Their users won't automatically get access to the replacement, and old links will stop working. The hands-on test only covers a simple case. How well complex GPTs with external integrations migrate remains an open question. For companies, the switch to plugins with MCP integration is more than a rebrand. It allows deeper integration but also demands more effort. If you're affected, check the notification in your account and, before December 11, use fixed test inputs to compare whether the replacement really does what your old GPT could.