Tracking enterprise infrastructure's evolution in the AI era

News

Users Decry Problems in OpenAI-Mandated Custom GPT-to-Plugin Migration

• By David Ramel
• 09/30/2026

I have created a bunch of custom GPTs that I use with my company-paid OpenAI ChatGPT account and have just begun the mandatory migration of the useful GPTs to a plugin system, which we detailed earlier.

As we reported, OpenAI has set deadlines for the conversion.

The migration process has been frustrating for me, as the custom GPT functionality I rely on didn't survive the migration to plugins. The new plugin system imposes limitations that were not present in the previous setup, leading to a loss of productivity and increased complexity in maintaining my workflows.

For example, I had a custom GPT that helps me write initial article drafts, doing the heavy grunt work of research and structuring content, after which I fact-check things and do the actual human reporting. To help with fact checking, the GPT generated an "Evidence Pack" that provided links and explanations to help me verify the initial draft content. After the migration to the plugin system, the evidence pack and some other functionality disappeared, forcing me to manually perform tasks that were previously automated.

I complained to Chat GPT, and its solution was a long initial prompt that basically repeats some of the instructions in my custom GPT. I wasn't happy with that solution, as I preferred the sparse prompts used with my custom GPT. I find that I'm not alone, as there is much online backlash to the forced migration, much of it mirroring my experience closely.

For example, one user in an OpenAI Developer Community discussion reported a familiar problem: "In my case instructions and knowledge files seem to be completely ignored so it's basically running without instructions." The user reported better results after manually adding an instruction to follow the relevant skill, even after invoking the plugin with an @ mention. Another participant described responses drifting back toward ordinary ChatGPT behavior after several messages, despite explicitly selecting the plugin.

Missing content also figured in a Sept. 28 complaint about template adherence. A user who relied on custom GPTs for clinical documentation reported that migrated plugins changed headings, reorganized sections, omitted information and introduced details not supplied for the current client. The user said they tested the same input with the original GPT and its replacement and provided OpenAI Support with a synthetic template and both outputs. Those comparison materials were not included in the public post.

Fixing such problems can introduce another hurdle. In a Sept. 29 bug report, a user said migration completed without reporting problems, but testing revealed inconsistent enforcement of an existing instruction. When the user tried to add a corrective rule through Edit plugin, Plugin Creator could read the instructions but could not publish the update because it lacked the plugin's backend identifier. The report included reproduction steps and a screenshot; another participant reported the same error.

OpenAI itself acknowledges that transferring a configuration does not guarantee equivalent results. Its migration FAQ states, "A migrated plugin may respond differently." The company tells users to compare familiar prompts and check whether the replacement follows instructions, uses reference material and produces complete output in the required format. For users whose custom GPTs already handled those tasks, that means testing and repairing workflows they had previously configured.

Other complaints extend beyond the quality of the generated answers:

• Editing friction: In a Sept. 28 post, one creator described straightforward configuration changes becoming drawn-out exchanges with an AI assistant. "Previously, if I needed a quick one-line fix, I would just open the editor and change it." The user also reported slower operation and differences between Chat and Work modes.
• Blocked employee access: A ChatGPT Business workspace owner reported that a migrated human-resources assistant remained private, with its sharing control disabled. Employees could still access the original GPT, but could not use its replacement. See below for more on this.
• Disrupted customer distribution: In a discussion about access after migration, creators who distributed GPTs to paying customers through links described difficulty preserving that arrangement. One reported operating 47 GPTs for 25 active customers, while another emphasized that the conversion itself worked but customer access remained unresolved. Both described asking support for an equivalent way to share privately with external customers.
• Usage consumed by repairs: In that same discussion, a user reported exhausting most of both a five-hour allowance and a weekly allowance while trying to restore one plugin's previous functionality. The user also described difficulty managing reference files through conversational editing. Those usage figures were self-reported, but the complaint highlights another potential cost of rebuilding a previously working setup.
• Lost assistant identification: A Sept. 28 feature-regression report described difficulty identifying which GPT or plugin had produced an older conversation, particularly after moving chats into Projects. The author argued that missing assistant identities and visual cues made historical work harder to interpret and requested persistent records of the originating assistant and model.
• Educational workflows needing reconstruction: In a public letter to OpenAI, an educational creator described approximately 155 specialized GPTs connected to textbook exercises through direct links. The concern was that requiring learners to locate and invoke replacement plugins would complicate a previously straightforward process and require revisions to instructional links, documentation and training.
• Client workflows behaving differently: In a Sept. 29 Reddit discussion about migrating client GPTs, an agency owner asked how others were handling the transition. One commenter reported that migrated assistants failed to follow instructions and performed actions their instructions explicitly prohibited. Other participants discussed trying Projects as an alternative, with mixed reports.
• Reduced configuration control: In another Reddit migration discussion, users described poorer results with uploaded templates and difficulty inspecting or editing their configurations. One reported that a conversation starter was truncated during migration and remained incomplete after an attempted repair through Plugin Creator.

So far, OpenAI hasn't mounted much of a response to the user complaints, one exception concerns the sharing complaint in the list above.

In that discussion, the OpenAI_Support account responded on Sept. 29, signing as Ziad, by reiterating that migrated plugins start private and do not inherit the original GPT's sharing settings. The reply directed the workspace owner to plugin-sharing controls and workspace permissions, asking whether those options had become available.

Another administrator replied that the described plugin permissions were missing from their workspace and that migrated GPTs still could not be shared internally. As of this article's Sept. 30 review, the thread showed no confirmed resolution.

The deadline for migration is Dec. 11, 2026, although qualifying Enterprise workspaces with an approved deferral have until Feb. 11, 2027.