OpenAI Is Retiring Custom GPTs. The Bigger Story Is What Comes Next
Updated: 12 hours ago
September 15, 2026 | Report | OpenAI Retiring Custom GPTs
OpenAI once envisioned millions of specialized versions of ChatGPT and built a marketplace around them. Now it is retiring Custom GPTs in favor of plugins, reusable skills and connected applications.
The change is more than product housekeeping: it represents a fundamental shift in how OpenAI thinks artificial intelligence should be customized.
When OpenAI introduced GPTs on November 6, 2023, the pitch was unusually democratic for a technology platform.
Anyone could build a specialized version of ChatGPT.
A teacher could create a mathematics tutor. An investment firm could configure an internal research assistant. A company could package its policies and documentation into an employee-support tool. A consultant could encode a methodology into a chatbot and distribute it to clients.
Programming was optional.
OpenAI described GPTs as customized versions of ChatGPT that could combine instructions, additional knowledge and capabilities for a particular purpose. The company explicitly positioned them as a way for ordinary users—not simply software developers—to shape how artificial intelligence behaved.
The idea spread rapidly.
By January 10, 2024, only about two months after the initial announcement, OpenAI said users had created more than 3 million Custom GPTs. It launched the GPT Store that day, providing searchable categories, rankings and featured GPTs while promoting the possibility that builders could eventually earn money based on usage.
Custom GPTs appeared to be developing into an AI-era equivalent of the app store.
Now OpenAI is dismantling that architecture.
According to a new retirement and migration FAQ published in September 2026, OpenAI plans to retire Custom GPTs across ChatGPT and is encouraging creators to migrate their workflows to a newer system built around plugins, skills and connected apps.
For affected Enterprise workspaces, OpenAI currently lists four major milestones:
September 11, 2026 — administrators were notified of the transition.
September 17, 2026 — OpenAI is targeting the release of its migration experience and user notification banner.
September 25, 2026 — creation of new Custom GPTs is scheduled to end in affected Enterprise environments.
December 11, 2026 — existing Custom GPTs are scheduled to stop running.
OpenAI stresses that these dates remain subject to change and that account- and plan-specific availability may differ. The broader transition applies across ChatGPT plans, but the detailed schedule published so far is centered on Enterprise.
This Is Not the Retirement of Customization
The easiest interpretation is that OpenAI tried Custom GPTs, decided the product did not work and is killing it.
The evidence does not support such a simple conclusion.
In fact, OpenAI's own enterprise data suggest precisely the opposite.
Its 2025 State of Enterprise AI report said weekly users of Custom GPTs and Projects increased approximately 19-fold year-to-date. In recent months covered by the report, about 20% of all Enterprise ChatGPT messages were being processed through a Custom GPT or Project.
OpenAI also highlighted Spanish banking giant BBVA, which it said was regularly using more than 4,000 GPTs across the organization.
Those numbers are significant.
They suggest OpenAI is not retiring a feature that companies simply ignored. It is replacing an architecture that had already become embedded in at least some enterprise workflows.
That makes the decision far more consequential.
What Exactly Was a Custom GPT?
Despite their name, Custom GPTs were not independently trained artificial-intelligence models.
Creating one did not mean training a new GPT-4 or GPT-5 from scratch.
Instead, users constructed a specialized interface around OpenAI's underlying models.
A Custom GPT could contain several layers.
Instructions established persistent behavior: how the GPT should respond, what methodology it should follow, what tone to use and which rules should govern its decisions.
Knowledge files allowed users to upload documents that the GPT could reference when responding.
Conversation starters provided predefined examples to help users understand what the assistant could do.
Capabilities could include functions such as web search, image generation and data analysis.
Apps could connect the GPT with supported external services.
And Custom Actions allowed more sophisticated GPTs to communicate with outside APIs.
OpenAI's current GPT documentation says a GPT can combine instructions, knowledge and capabilities to produce a specialized version of ChatGPT.
Actions were particularly important for advanced deployments.
A company could provide an OpenAPI schema describing an external service, configure authentication through an API key or OAuth, and allow the GPT to retrieve information or trigger operations outside ChatGPT.
The result was something between a prompt template, a knowledge base, an application and a lightweight AI agent.
Why People Used Them
The attraction was not that Custom GPTs contained fundamentally different intelligence.
It was that they could preserve expertise.
Consider an analyst who produces investment reports.
Without a Custom GPT, that analyst might repeatedly explain:
how the report should be organized;
which valuation metrics should be calculated;
how risks should be categorized;
how sources should be cited;
how much detail each section requires;
which writing style should be used.
A Custom GPT could preserve those instructions.
The same concept applied inside corporations.
Human-resources teams could package company policy.
Sales departments could encode proposal methodology.
Engineering teams could build onboarding assistants.
Consulting firms could preserve internal frameworks.
Compliance departments could create repeatable review processes.
Rather than teaching ChatGPT the same procedure every time, organizations could institutionalize the procedure.
That was why the 19-fold increase in enterprise usage cited by OpenAI mattered: workers were moving from occasional prompting toward repeatable AI workflows.
The Problem: Every Workflow Became Another Bot
The strength of Custom GPTs was also their architectural weakness.
Imagine a large corporation that creates:
a procurement GPT;
a legal-review GPT;
an earnings-analysis GPT;
a customer-support GPT;
a marketing GPT;
an HR GPT;
a cybersecurity GPT;
and dozens—or thousands—more.
Every GPT potentially contains separate instructions, reference files, permissions and integrations.
Useful capabilities become trapped inside individual assistants.
Suppose the legal-review GPT contains an excellent procedure for extracting contractual obligations. The procurement GPT may also need that capability.
Under a bot-centric architecture, the organization may duplicate the instructions or create another specialized GPT.
The larger the ecosystem becomes, the greater the risk of fragmentation.
OpenAI is now moving toward a different model.
From Custom GPTs to Plugins
The central replacement is the plugin.
In OpenAI's current architecture, a plugin can package multiple types of functionality together.
A plugin may contain:
skills providing instructions and workflow logic;
connected apps that access outside services, accounts and data;
and administrative templates that help organizations configure those applications.
OpenAI describes plugins as a way for ChatGPT and Codex to complete workflows by combining reusable instructions, connected tools and other capabilities.
That sounds superficially similar to Custom GPTs.
Architecturally, however, there is an important difference.
The specialization no longer has to be inseparable from a particular chatbot.
Skills May Be the More Important Technology
Inside the plugin system sits another increasingly important component: skills.
OpenAI describes skills as reusable and shareable workflows that help ChatGPT perform specific tasks more consistently.
A skill can contain:
instructions;
examples;
supporting resources;
reusable steps;
scripts;
and code.
Once installed, ChatGPT can automatically select one or more appropriate skills when they are useful for a particular request.
That seemingly small change represents a different philosophy of AI software.
Under the Custom GPT model:
User chooses specialized assistant → specialized assistant performs task.
Under the emerging skills architecture:
User asks general AI → AI identifies task → relevant specialized capability is loaded.
The burden of choosing the correct specialist gradually moves from the human to the AI.
What Actually Migrates
For users who have invested significant time building Custom GPTs, migration is not simply a rename operation.
OpenAI says that during the planned migration process, a GPT's instructions become a skill inside the replacement plugin.
Connected apps can migrate into the plugin as apps.
But several other elements require review.
OpenAI warns that:
conversation starters may not transfer;
previous conversations may not transfer;
the GPT's selected model does not carry over;
reference material and templates should be reviewed;
and migrated workflows may behave differently from the original GPT.
That final warning is important.
A migration may preserve the instructions without perfectly preserving the behavior.
Custom Actions Are the Biggest Migration Problem
The most consequential complication involves Custom Actions.
They do not automatically migrate.
Organizations that connected a GPT to an outside API will need to evaluate those integrations separately and potentially rebuild them using supported connectors or a custom Model Context Protocol server.
OpenAI explicitly cautions users not to assume that a rebuilt integration will reproduce every capability of the original action.
For casual GPT creators, migration may therefore be little more than transferring instructions and knowledge.
For a company that built a Custom GPT into operational infrastructure, it may resemble a genuine software migration project.
The Model Context Protocol Enters the Picture
MCP—Model Context Protocol—is becoming increasingly important across the AI industry because it provides a standardized way for artificial-intelligence systems to interact with external tools and information sources.
OpenAI's decision to steer more complex integrations toward connectors and MCP reflects a larger industry movement away from proprietary one-off integrations.
Microsoft's current Copilot Studio infrastructure, for example, also supports MCP tools that connect agents to custom services, databases and APIs.
This suggests that the future competitive battleground may not be which AI company has the largest catalog of isolated assistants.
It may instead be which platform can most effectively orchestrate a common ecosystem of skills, data and tools.
OpenAI's Official Explanation—and What It Does Not Say
OpenAI's retirement FAQ is relatively direct about what it wants users to do.
It says:
“We're planning to retire custom GPTs, and encourage you to consider using plugins as replacements.”
The company explains that plugins can combine reusable instructions with connected applications so guidance can operate alongside the tools and information required by a task.
What OpenAI does not provide is a detailed postmortem.
It does not say Custom GPT adoption was weak.
It does not say the GPT Store failed economically.
It does not say maintaining GPTs became prohibitively expensive.
And it does not explicitly identify fragmentation as the reason for retirement.
Those would be interpretations rather than established facts.
The strongest evidence for OpenAI's reasoning therefore comes from examining the architecture it is building.
OpenAI Is Separating Intelligence From Expertise
The Custom GPT model packaged several things together:
AI model;
persistent instructions;
knowledge;
tools;
external integrations;
identity;
distribution.
The new architecture begins separating them.
The foundation model provides general intelligence.
Skills provide repeatable procedural expertise.
Apps provide data and actions.
Plugins combine capabilities into workflows.
ChatGPT or Codex becomes the orchestration layer.
That is arguably a more durable architecture.
A procedure such as “prepare our weekly investment-committee report according to these rules” does not necessarily need its own chatbot.
It can instead exist as a skill that whichever capable OpenAI model is operating can invoke.
Anthropic Has Reached a Very Similar Conclusion
OpenAI is not alone.
Anthropic introduced Agent Skills for Claude in October 2025.
The company describes Skills as organized packages of instructions, scripts and resources that Claude can discover and load dynamically.
Anthropic's explanation is remarkably close to the logic underlying OpenAI's transition.
Rather than building what Anthropic describes as fragmented, custom-designed agents for individual use cases, Skills let organizations package procedural expertise into reusable components.
Anthropic later published Agent Skills as an open standard for cross-platform portability.
That provides meaningful validation for OpenAI's architectural direction.
Two major AI laboratories independently appear to have concluded that reusable capabilities are a better long-term abstraction than building a separate AI persona for every workflow.
But Google Is Doing Something Different
Google provides perhaps the strongest argument against completely retiring the Custom GPT concept.
Google's Gems remain customized versions of Gemini.
Users can give a Gem persistent instructions, add knowledge files and create specialized assistants for recurring activities.
And Google has continued expanding the model.
In September 2025, Google introduced the ability to share custom Gems with other people using access controls similar to Google Drive.
At the same time, however, Google is also building skills.
In 2026, Google began expanding Workspace Studio Skills, allowing organizations to convert reusable prompts, rules, reference documents and templates into specialized capabilities that Gemini can invoke inside Workspace.
That means Google's strategy is not necessarily:
Custom assistants or skills.
It is:
Custom assistants and skills.
That difference deserves attention.
Google May Be Preserving Something OpenAI Is Discarding
There is a user-experience advantage to saying:
“Open the Due Diligence Gem.”
or:
“Use the Marketing Assistant.”
A named assistant has an identity.
Employees know what it does.
Managers know why it exists.
Teams can share it.
A business can assign responsibility for maintaining it.
The underlying technical implementation can still use modular skills.
Google's approach therefore raises a legitimate question:
Did OpenAI actually need to retire the Custom GPT interface, or could it have retained GPTs as a simple user-facing shell over its new plugin-and-skill architecture?
The answer is not obvious.
Microsoft Still Believes in Named Agents
Microsoft is taking an even more explicit agent-centric approach.
Copilot Studio continues to allow organizations to create custom agents.
According to Microsoft's current documentation, a custom agent can contain:
specialized instructions;
knowledge;
tools;
suggested prompts;
external integrations;
and custom behavior.
Those agents can then be published into Microsoft 365 Copilot, Teams and other environments.
Microsoft even maintains an Agent Store architecture in which organizational agents can be built, submitted for review and distributed to users.
Microsoft is simultaneously making those agents more modular by supporting MCP, workflows and external tools.
Its strategy suggests that visible agents and composable infrastructure do not have to be mutually exclusive.
The Competitive Divide
The major AI companies increasingly agree about the infrastructure.
They disagree more about the interface.
Company | Named Custom Assistant | Reusable Skills | External Tools / Apps | Current Direction |
OpenAI | Custom GPTs being retired | Yes | Plugins, apps, MCP | General AI + composable capabilities |
Anthropic | Less emphasis on standalone custom personas | Yes, Agent Skills | MCP and tools | General agents + portable skills |
Yes, Gemini Gems | Yes, Workspace Skills | Workspace integrations | Hybrid model | |
Microsoft | Yes, custom Copilot agents | Capability modularization | Tools, workflows, MCP | Agent-centric enterprise platform |
Sources: OpenAI product documentation; Anthropic Agent Skills documentation; Google Gemini and Workspace documentation; Microsoft Copilot Studio documentation.
The Case for OpenAI's Decision
From an engineering perspective, OpenAI's strategy has considerable logic.
First, skills are reusable.
A procedure does not have to be duplicated across multiple assistants.
Second, capabilities can become model-independent.
As OpenAI upgrades its underlying models, organizations do not necessarily need to rebuild an entire customized chatbot.
Third, ChatGPT can choose skills automatically.
OpenAI says installed skills can be selected when their descriptions match the user's request.
The long-term vision is compelling: Users should not need to understand the software architecture before asking for work to be done.
The AI should recognize the task and assemble the appropriate capability itself.
The Case Against It
The strongest criticism is not technical.
It is about product design.
Custom GPTs were easy to understand.
Build a specialist.
Name it.
Give it instructions.
Give it documents.
Share it.
That metaphor makes sense to almost anyone.
“Plugins containing skills connected to apps, some of which can automatically activate based on the task” may be architecturally better while being cognitively harder for ordinary users to understand.
OpenAI therefore risks replacing a technically imperfect abstraction with a more sophisticated abstraction that requires more explanation.
The GPT Store Question
The retirement also raises questions about one of OpenAI's most ambitious early platform strategies.
When OpenAI launched the GPT Store in January 2024, it presented the marketplace as a place where people could discover popular custom assistants across areas such as writing, research, education and programming.
More than three million GPTs had already been built.
The Store looked increasingly like a potential AI equivalent of Apple's App Store.
That vision now appears to have changed.
OpenAI's newer Plugin Directory is much closer to a marketplace for capabilities and integrations than a marketplace exclusively for independent AI personalities.
The difference is subtle but profound.
The unit of distribution is moving from:
“Here is another AI.”
toward:
“Here is something your AI can do.”
Figure 1 — OpenAI's Planned Custom GPT Migration Runway
Enterprise dates currently published by OpenAI; milestones remain subject to change.
Why December 11 Should Not Be Treated as the Migration Date
For organizations with meaningful Custom GPT deployments, December 11 should be considered a shutdown deadline, not a migration date.
OpenAI recommends identifying:
which GPTs are still actively used;
who owns them;
who needs access;
which reference files they rely on;
which integrations they use;
and which prompts can be used to compare old and new behavior.
Migrated replacements should then be tested using both familiar prompts and difficult edge cases.
OpenAI specifically recommends confirming that the replacement:
selects the correct skill;
follows the intended instructions;
uses the expected reference material;
produces required outputs and files;
and still possesses the necessary integrations.
For businesses, this should be treated as governance work—not simply an interface update.
The Most Interesting Contradiction
Perhaps the most revealing data point in this entire story is OpenAI's own 2025 enterprise research.
Custom GPTs and Projects were not disappearing from enterprise use.
They were growing.
Weekly users had increased approximately 19× during the year, and around one-fifth of Enterprise messages were running through Custom GPTs or Projects.
That changes how the retirement should be interpreted.
This is not necessarily OpenAI abandoning something nobody wanted.
It may be OpenAI deciding that something users demonstrably wanted was built around the wrong long-term abstraction.
Is OpenAI Making the Right Decision?
My assessment is that OpenAI is probably making the right architectural decision and a more debatable product decision.
The architectural evidence is strong.
Anthropic independently reached a very similar conclusion about composable skills.
Google is introducing reusable skills throughout Workspace.
Microsoft is expanding standardized tools and MCP connectivity.
Across the AI industry, procedural expertise is becoming modular.
That trend is real.
The future AI interface probably does contain fewer situations in which a human must manually decide which narrow chatbot to use before asking a question.
Eventually, a sufficiently capable general-purpose system should understand the task, retrieve the correct knowledge, invoke the appropriate skill, access the necessary applications and complete the workflow.
But OpenAI may be moving too aggressively in eliminating the visible Custom GPT layer.
Google's model is instructive.
A user can still create and share a named Gem while Google builds skills underneath its broader enterprise architecture.
Microsoft likewise retains recognizable agents while making the underlying infrastructure increasingly modular.
Those approaches preserve something valuable:
identity.
A company may genuinely want an “Investment Committee Assistant.”
A newsroom may want an “Investigations Editor.”
A law firm may want a “Contract Review Agent.”
Those names provide context and organizational meaning.
There is no technical reason reusable skills and named assistants must always be mutually exclusive.
The Larger Shift
The deeper story is not really the death of Custom GPTs.
It is the gradual disappearance of the chatbot as the fundamental unit of AI software.
The first generation of generative AI asked users to choose an application.
The second generation asked them to choose a model.
The Custom GPT era asked them to choose a specialized AI assistant.
The emerging agent era increasingly asks only one thing:
What do you want done?
Everything underneath—the model, skill, tool, application, database, code execution environment or workflow—can increasingly be selected dynamically.
That is the architecture toward which OpenAI appears to be moving.
If the company succeeds, Custom GPTs may eventually be remembered as an important transitional technology: the bridge between a world of individual AI chatbots and a world in which a general-purpose AI system dynamically assembles whatever expertise it needs.
If OpenAI fails, the criticism will be equally straightforward.
It will have taken one of the easiest AI customization concepts ever created, replaced it with a more complicated vocabulary of plugins, skills and apps, and surrendered the simple custom-assistant experience to Google and Microsoft.
The engineering case is persuasive.
The product case still has to be proven.
Primary Sources
OpenAI — Custom GPT retirement and migration FAQ. The principal source for the September–December 2026 transition, migration mechanics, custom-action limitations and access changes.
OpenAI — Introducing GPTs, November 6, 2023. Original announcement explaining the philosophy, features and intended ecosystem behind Custom GPTs.
OpenAI — Introducing the GPT Store, January 10, 2024. Source for the more-than-three-million-GPT figure and the original marketplace strategy.
OpenAI — State of Enterprise AI 2025. Source for the approximately 19× increase in weekly Custom GPT/Project users, the roughly 20% share of Enterprise messages and BBVA's use of more than 4,000 GPTs.
OpenAI — Skills in ChatGPT. Documentation covering reusable workflow instructions, scripts, examples, code and automatic skill selection.
OpenAI — Plugins in ChatGPT and Codex. Documentation explaining how plugins combine skills, applications and other capabilities.
Anthropic — Equipping Agents for the Real World with Agent Skills. Anthropic's explanation of composable skills and its argument against fragmented custom agents.
Google — Gemini Gems documentation and sharing announcement. Sources for Google's continued support for custom Gemini assistants.
Google Workspace — Workspace Studio Skills. Source for Google's parallel strategy of introducing reusable skills alongside Gems.
Microsoft — Copilot Studio agent documentation. Source for Microsoft's continued investment in user-visible custom agents, knowledge and tools.
Comments