TL;DR
A custom GPT is a version of ChatGPT set up for one job, with its own instructions, knowledge files, capabilities and optional actions. Building one does not train a new model. OpenAI retires custom GPTs on Dec 11, 2026, or on Feb 11, 2027 for FedRAMP and approved-deferral workspaces. Personal Free, Go, Plus and Pro accounts can no longer create new GPTs, although existing ones stay usable and editable. Where a managed workspace still allows creation, you build one in GPT Builder in eight steps. Before the deadline, migrate each GPT to a plugin and rebuild any custom actions, because those do not carry over.
Key Takeaways A custom GPT is ChatGPT configured with instructions, knowledge, capabilities and either apps or actions, with no retraining of the model. New GPT creation is closed on personal accounts, and affected Enterprise workspaces stopped creating new GPTs on Oct 26, 2026. Custom GPTs retire on Dec 11, 2026, when GPT pages go dark, but past conversations stay readable. Instructions, knowledge files and connected apps carry into a plugin, while custom actions and the selected model do not. Migration copies the latest published version only, so publish your final edits before you press Migrate to plugin. When users sit outside ChatGPT or need record-level permissions, a Copilot Studio agent or an API-built app is the better home. Watch on YouTube
GPT-5.6 Is Here. Is Your Data Ready?
A Kanerika video on getting enterprise data ready for new GPT models, the same groundwork a custom GPT’s knowledge files depend on.
Can You Still Build a Custom GPT in November 2026?
November 6, 2026. OpenAI’s help center now carries one plain line. “Custom GPTs are scheduled to retire on Dec 11, 2026.”
Any HR helper, sales coach or IT policy bot built in ChatGPT must move before Dec 11, 2026 on the standard schedule. OpenAI’s retirement and migration FAQ also explains where those assistants go next.
The change has not reached most how-to guides yet. Most of the top-ranking pages we reviewed for this keyword still skip the retirement. They walk readers through a build that no personal account can now complete.
So the honest answer is yes, but only in some places, and only for a few more weeks. The parts you configure in a GPT still matter, though, because they become the parts of its successor.
Who Can Create, Edit and Use a Custom GPT Today
Access now depends, above all, on the kind of account you hold. According to OpenAI’s guide to creating and editing GPTs , the rules break down like this.
Personal accounts (Free, Go, Plus, Pro): new GPT creation and publishing are switched off. Existing GPTs stay usable, and you can still edit one if your plan and permissions allow it.Business, Enterprise and Edu workspaces: creation, editing and publishing depend on product availability, workspace settings and the permissions your admin grants.Affected Enterprise workspaces: new GPT creation ended on Oct 26, 2026, except in FedRAMP workspaces. Edits to existing, unmigrated GPTs are still allowed.Mobile apps: you can use a GPT on iOS or Android, but building and editing happen on the web only.For brand-new reusable workflows, OpenAI now points people to Plugin Creator instead. In practice, that means most readers of this guide fall into one of two groups. Either they still have creation rights in a managed workspace, or they own GPTs that need a careful move before the deadline.
The Custom GPT Retirement Timeline
The dates below come straight from OpenAI’s retirement FAQ. Enterprise workspaces saw the earliest steps, while personal accounts mainly feel the final date.
Date What happens Who it affects Sep 11, 2026 Retirement notice sent to admins Enterprise workspace admins Sep 22, 2026 Target date for the migration experience and the user banner Enterprise workspaces Oct 1, 2026 Banner target for workspaces with a delayed banner Enterprise workspaces on the delayed track Oct 26, 2026 New GPT creation ends; edits to existing, unmigrated GPTs continue Affected Enterprise workspaces, except FedRAMP Dec 11, 2026 Standard retirement; custom GPTs and their GPT pages become inaccessible Every account without a later date Feb 11, 2027 Retirement for FedRAMP workspaces and approved deferrals FedRAMP and qualifying Enterprise workspaces only
One detail eases the pressure. Your existing conversations with a GPT remain accessible after retirement, so chat history is safe. The GPT itself, its page and its configuration are what disappear.
What Is a Custom GPT and How Does It Work?
A custom GPT is a version of ChatGPT configured for a specific purpose. OpenAI’s own GPTs in ChatGPT overview describes it as a mix of instructions, knowledge and selected capabilities. Nothing gets retrained. The builder wraps one of the existing OpenAI GPT models in a standing brief that every conversation then starts from.
That brief has limits worth knowing early. A GPT does not use your saved memory, your account-level custom instructions or your previous conversations, so each chat starts fresh. Also, the “instructions” field inside a GPT is a separate thing from ChatGPT’s personal Custom Instructions setting, even though the names sound alike.
The Building Blocks of a Custom GPT
The builder has six fields, and four of them do the heavy lifting: instructions, knowledge, capabilities and actions. Each part also has a counterpart in the plugin that replaces it, so the list below notes where each one lands.
Instructions: the standing brief on role, scope, tone and output format. In a plugin, these become a skill.Conversation starters: suggested first prompts that show users what the GPT is for.Knowledge: uploaded files the GPT can search for answers. These copy into a plugin’s reference files.Capabilities: web search, image generation, canvas, and Code Interpreter and Data Analysis, switched on per GPT.Apps: connected apps that bring outside tools into the chat. They move into the plugin as apps.Actions: calls to external APIs defined by a schema. These do not transfer, so they need a rebuild.One rule shapes the design from the start. A GPT can use either apps or actions, but not both at the same time. So a team that needs both a ticketing system and a document store must pick one route. Otherwise, the work has to move somewhere with more room.
Custom GPT vs ChatGPT, Projects, Fine-Tuning, RAG and the API
A GPT sits in the middle of a range of options. Some change less than it does, while others reach further into the model or the product. The table compares them on what each one actually changes.
Option What it changes Where it runs Typical use Main limit Regular ChatGPT chat Nothing beyond your account settings ChatGPT One-off questions and drafts Context must be re-supplied every time ChatGPT Project Groups your own chats and files around ongoing work ChatGPT A personal or small-team workstream Built for your own work, not a shared tool Custom GPT Instructions, knowledge, capabilities, apps or actions ChatGPT only A repeatable task many users run the same way Retires Dec 11, 2026; no memory across chats; cannot be embedded Fine-tuning Model weights, through training examples OpenAI API Consistent style or format at high volume Needs training data and evaluation; poor at fresh facts Engineered RAG A retrieval pipeline over your indexed content Your own app or platform Large, changing document sets with access control Engineering and upkeep API-built app Everything, in code Your product, website or internal tools Assistants inside a product or customer site Build and running cost
The comparison readers ask about most, though, is the website one. A GPT cannot be embedded on an external site, and OpenAI says plainly that product assistants belong on the API instead. For the trade-off between training and retrieval, Kanerika’s guide to RAG vs fine-tuning goes deeper, and the primer on retrieval-augmented generation explains the pipeline itself.
Is There a Claude or Gemini Equivalent of a Custom GPT?
Yes, broadly. Anthropic’s Claude Projects and Google’s Gems both let you save instructions and reference material, then reuse them across chats. Their feature sets differ and change often, though, so compare against each vendor’s current documentation before you commit a team to one.
How to Create a Custom GPT in GPT Builder
These steps apply where creation is still enabled, which today means eligible managed workspaces and FedRAMP workspaces. They also cover editing a GPT you already own. And because the same fields carry into a plugin, the thinking behind each step outlives the builder itself. The eight steps group into five moves: define the job, write instructions, add approved knowledge, test, and share with an owner.
The running example is an internal IT policy assistant. It answers staff questions about laptops, passwords and software from approved policy documents. Then it points people to the service desk when a question goes beyond policy.
Step 1: Define One Job, One Audience and Its Boundaries
Write one sentence that names the job, the users and the line the GPT must not cross. For example, “Answer IT policy questions from the approved handbook, and send access changes to the service desk.” A narrow brief like this makes every later step easier to test.
Step 2: Open the Builder and Set Name, Description and Starters
Go to Explore GPTs, or chatgpt.com/gpts, and then select Create. The builder has two modes. The conversational builder interviews you and drafts the setup, while the Configuration view lets you fill each field directly.
Most teams start in the conversational mode, then switch to Configuration to tighten the wording.
After that, give the GPT a plain name, a one-line description of its job and three or four conversation starters. Good starters double as a menu, so a user who opens “IT Policy Helper” sees questions like “How do I request new software?” right away.
Step 3: Write the Instructions
Instructions carry most of the behavior, so they get their own section below. At this stage, write the role, the scope, the sources the GPT should trust and what it should do when the answer is missing. Then keep the draft short enough that a colleague could review it in two minutes.
Step 4: Upload Approved Knowledge Files
Attach only documents that someone owns and has approved. A GPT accepts up to 20 files, and each file can also be up to 512 MB. For the IT example, that means the current handbook, the software catalog and the password policy, not last year’s drafts.
Step 5: Turn On Only the Capabilities the Job Needs
The capability switches cover web search, image generation, canvas, Code Interpreter and Data Analysis, and apps. A policy assistant needs none of them except perhaps web search, so leave the rest off. Fewer capabilities mean fewer surprises, and they also make test results easier to read. You can also set a recommended model here.
Step 6: Add Actions If the GPT Must Call an API
Add an action only when the GPT has to read from or write to another system, such as checking a ticket’s status. Each action also needs an API schema and an authentication method. Start with read-only calls, because a wrong answer is easier to fix than a wrong write.
Step 7: Test in Preview Against a Fixed Question Set
The Preview pane lets you chat with the GPT while you edit it.
Use that pane against a fixed set of 20 to 30 real questions, including a few that should be refused or escalated. Keep the set and rerun it after every change. For larger sets, a second model can score the answers, as Kanerika’s guide to LLM-as-a-judge evaluation explains. The same discipline also appears in AI agent evaluation .
Step 8: Save, Share and Assign an Owner
Finally, save the GPT, choose who can see it and name one owner who answers for its content. Version history sits in the ••• menu, so you can roll back a bad edit. Publishing also matters more than it used to, because migration copies only the latest published version.
Checklist
Generative AI Checklist
A practical checklist for secure generative AI adoption, covering use-case fit, data handling, access controls and governance before an assistant reaches users.
Get the Checklist → How to Write Custom GPT Instructions That Hold Up
Good instructions read like a short operating manual, not a personality sketch. OpenAI’s builder guidance favors explicit steps, such as “When X happens, do Y”, and positive instructions over long lists of prohibitions. It also suggests brief examples of acceptable and unacceptable answers, laid out under headings.
An Instruction Checklist
Before you publish, check that the instructions cover each of these points. The broader craft is covered in Kanerika’s prompt engineering best practices .
Role and audience: who the GPT serves and in what voice.Scope: the questions it answers and the ones it hands off.Sources and precedence: which file wins when two documents disagree.Tool use: when to search the web or call an action, and when not to.Output format: length, structure and citation style.Missing information: what to say when the files have no answer.Sensitive data: what users should never paste in.Escalation: the exact route to a human.Example: Instructions for an Internal IT Policy GPT
Here is a compact version for the running example. Treat it as a pattern to adapt rather than a finished policy, because every workspace words its rules differently.
ROLE
You are IT Policy Helper for employees of Contoso.
SCOPE
Answer questions about devices, passwords, software requests
and acceptable use, using only the attached policy files.
SOURCES
1. IT_Handbook_2026.pdf wins over every other file.
2. If files disagree, say so and quote both sections.
WHEN THE ANSWER IS MISSING
Say "The policy files do not cover this." and give the
service desk link. Do not guess.
ACCESS REQUESTS
For admin rights, VPN or new system access, explain the
request form and stop. Never approve or promise access.
FORMAT
Short answer first, then the policy section name and page.One caution also applies to any example like this. Instructions shape behavior, but they are not a security boundary, and a determined user can sometimes talk a model around them. Real limits instead belong in permissions and backend checks, as covered in Kanerika’s guide to AI guardrails . When the GPT migrates, this same text becomes the plugin’s skill, so the effort carries forward.
Preparing Knowledge Files for Reliable Answers
Knowledge files decide whether answers are right, so they deserve the same care as the instructions.
In general, text-forward documents work best. Complex layouts, scanned pages and dense multi-column tables are harder for the GPT to use. So convert them to clean text or simple tables first.
Three habits keep a knowledge set healthy. First, add a version date and an owner to each file, so stale content is easy to spot. Second, remove duplicates and superseded versions, because two conflicting policies produce confident, contradictory answers. Third, keep rules in the instructions and facts in the files.
Uploading a file also does not guarantee that every passage gets used in every answer. Test the questions that depend on small details, for example a specific limit buried on page 40. The wider discipline of AI knowledge management also applies here, and on migration these files copy straight into the plugin’s reference files.
Case Study
80% Ticket Auto-Resolution with LLM-Driven AI
Kanerika built a knowledge base and prepared historical tickets for an LLM-driven ticket response system that auto-resolves 80% of tickets.
Read the Case Study → Custom GPT Actions: Connecting to APIs Safely
Actions are where a GPT stops answering and starts doing. They are also the part that needs the most care, both for security and for the move to a plugin.
How Actions Work
An action describes an external REST API with an OpenAPI schema. ChatGPT then turns a user’s plain request into the JSON call that the API expects. OpenAI’s GPT Actions documentation gives two typical uses, retrieving data, such as querying a data warehouse, and taking action, such as filing a Jira ticket.
Authentication is also set per action in the editor. The authentication options are None, API key and OAuth, and the default is None.
So an unconfigured action talks to its API with no credentials at all. Use OAuth wherever the API acts on behalf of a specific user, and release read-only actions before any that write. When an action starts placing orders or moving money, the design questions look a lot like those in agentic commerce .
Security Risks in Actions and Knowledge
When a GPT uses apps or actions, relevant parts of the user’s input may be sent to the third-party service. OpenAI does not audit or control how those services use or store that data. So every action domain needs the same vendor review as any other data processor.
Prompt injection is the second risk. A retrieved document or an API response can carry hidden text that tries to redirect the model.
For that reason, treat both as untrusted input, enforce authorization on the server rather than in the prompt, and never put secrets in instructions or knowledge files. Kanerika’s piece on RAG security covers the document side of this problem in more depth.
Watch on YouTube
AI Security for Enterprises: Kanerika’s CISO on Prompt Injection, ISO 27001 & SOC 2
Kanerika’s CISO explains how prompt injection works and which controls keep enterprise AI assistants, and the data behind them, safe.
Why Actions Need the Most Migration Work
Custom actions do not transfer through the migration workflow. According to OpenAI, rebuilding the connection may require a custom MCP server and some technical setup. That makes actions the first item to schedule, not the last.
MCP is an open standard, and Kanerika’s explainer on the Model Context Protocol walks through how it works. If an off-the-shelf connector already exists, the plugin may simply use it as an app instead. Kanerika’s overview of apps in ChatGPT shows what is already available.
Sharing, Privacy and Governance for Business Custom GPTs
A GPT shared with a whole workspace becomes a small internal product. It needs an owner, rules for who can see it and a plan for the day it changes.
Sharing Options and the GPT Store
In managed workspaces, the sharing options may include specific people, the whole workspace, anyone with a link, or the GPT Store when permitted. Admins decide which of these methods are allowed. Personal accounts, as noted above, can no longer publish new GPTs at all, so the Store now matters mainly inside managed workspaces that permit it.
What Happens to Your Data
Data from Business, Enterprise and Edu plans is not used for model training by default. Consumer plans work the other way, so their data may be used for training unless the user opts out. On the question people search for most, builders cannot view the individual conversations users have with their GPTs.
That privacy line also shapes support. Because an owner cannot read users’ chats, feedback has to come through another channel, such as a short form or a team inbox named in the GPT’s description. Without one, the owner only hears about wrong answers when someone complains.
Governing a Workspace Full of GPTs
Workspaces that allowed open building often hold many GPTs with unclear owners. A short governance routine fixes that, and it doubles as migration prep.
Inventory: list every GPT, its owner, its audience and whether it uses actions. A full workspace inventory currently needs a workspace Owner.Naming: use a team prefix so users can tell official tools from personal experiments.Review dates: give each shared GPT a date when its files and instructions get checked.Action review: approve every external domain an action calls, as you would any vendor.Retirement: archive GPTs nobody owns instead of migrating them.This routine sits inside a wider generative AI governance program. It also surfaces assistants built outside approved channels, which Kanerika’s guide to shadow AI detection treats as a risk in its own right.
AI Assessment
How Ready Is Your Workspace for Governed AI Assistants?
Score your AI maturity across data, governance and security to see which GPTs are ready to migrate and which need more controls first.
Start Your AI Assessment → Custom GPT Use Cases by Business Function
The strongest GPTs do one narrow job for one group of people. The table pairs common examples with the guardrail each one needs and a metric worth tracking. More ideas by industry appear in Kanerika’s roundup of generative AI use cases .
Function Example GPT Guardrail Metric to track HR Benefits and leave policy helper Answers from the current handbook only; routes personal cases to HR Share of questions answered without an HR ticket Sales Messaging and objection coach Approved claims and pricing language only Rep usage per week and manager review scores Customer support Agent-assist reply drafter Drafts only; a human sends every reply Handle time and reply edit rate IT Runbook and policy assistant No access approvals; links to the service desk Deflected tickets and wrong-answer reports Finance Expense and travel policy checker Cites the policy section for every answer Policy exceptions caught before submission Legal Clause lookup over approved templates Finds and quotes clauses; never gives legal advice Time to locate a standard clause
Support is where these assistants most often outgrow a chat window. A drafting GPT helps agents, but routing, voice and ticket updates usually belong in a wider contact center automation setup. Kanerika saw this with a member support agent connected to knowledge bases and Zendesk. It resolves 65% of queries instantly and routes low-confidence cases to live staff.
Case Study
65% of Member Queries Resolved Instantly by an AI Support Agent
Kanerika built an AI support agent connected to knowledge bases and Zendesk that resolves 65% of member queries and routes low-confidence cases to people.
Read the Case Study → Migrating a Custom GPT to a Plugin Before Dec 11, 2026
OpenAI is helping creators move their GPTs into plugins. A plugin brings reusable instructions, reference files and connected apps together. According to the ChatGPT plugins documentation , plugins can include skills and MCP servers and work in both ChatGPT and Codex, across web, desktop and mobile.
What Carries Over and What Does Not
Most of a GPT moves across on its own, so the work is smaller than it looks. The gaps sit in actions, model choice and sharing, which is why those need a plan.
Custom GPT part Plugin counterpart Action needed Instructions A skill inside the plugin Review the wording once migrated Knowledge files Reference files Remove stale files before migrating Connected apps Apps in the plugin Confirm each app still connects Custom actions Not transferred Rebuild, possibly as a custom MCP server Selected model Not carried over Check results on the default model Sharing settings Not carried over; the plugin starts private Re-share with the original audience
In Enterprise workspaces, the workspace defaults apply in place of the old model choice. That can shift answer style slightly, so rerun the test set from Step 7 against the migrated plugin.
A Migration Checklist for GPT Owners
Work through these steps in order, one GPT at a time, because parallel migrations make comparison harder.
Confirm the GPT is still used and still owned. Archive it if not. Finish any edits, then publish, because migration uses the latest published version and drops drafts. Save the test questions and a few current answers for comparison. Go to My GPTs and select Migrate to plugin. Compare the plugin’s answers with the saved ones. Rebuild custom actions, through an existing app or a custom MCP server. Share the plugin again, since it starts private. Tell users where the new version lives. Publishing here does not mean public sharing, so a private GPT can be published for migration without exposing it. Once migrated, the original GPT turns read-only, and its creator cannot delete it.
Migrating Many GPTs Across an Enterprise Workspace
Large workspaces also have a bulk route. The ChatGPT Admin plugin runs bulk migration asynchronously. Starting the job does not confirm that every GPT moved, so admins should review each result and test it before users switch. Pulling the full inventory still needs a workspace Owner, though, so line that person up first. Then review the plugin sharing and publishing permissions before users go looking for their old tools.
When a Custom GPT Is Not Enough
Some GPTs should not move to a plugin at all, because they have outgrown the format. The table shows where each option reaches users and what it takes to build. The first row includes the plugin successor, since that is the same tool’s next form.
Option Where users reach it Build effort Integration and control Best fit Custom GPT, moving to a plugin ChatGPT; plugins also reach Codex Low Instructions, files and apps; actions rebuilt via MCP A repeatable task for ChatGPT users Copilot Studio agent Teams, Microsoft Copilot, websites, mobile apps and more Low-code Prebuilt and custom connectors to company data Microsoft 365 organizations Responses API app Your product, site or internal tools Engineering team Full control of prompts, tools and data Customer-facing or embedded assistants Fully custom agent Any channel Highest Own orchestration, permissions and monitoring Regulated, multi-system workflows
Microsoft describes Copilot Studio as a graphical, low-code studio for building agents that connect to company data and publish to Teams and other channels.
For that route, Kanerika’s guide to Copilot Studio covers the build side, and Microsoft Copilot vs ChatGPT helps with the platform choice. On the developer side, the Assistants API was sunset on August 26, 2026 . New builds therefore go on the Responses API, covered in Kanerika’s OpenAI API guide .
Four Signals a Custom GPT Is the Wrong Tool
A GPT that works well can still hit a ceiling. These four signals mark where a GPT is the wrong tool to build, and where a plugin migration only postpones a bigger rebuild.
External deployment: customers, partners or a mobile app need the assistant, and they do not have ChatGPT seats.Granular permissions: two users asking the same question should get different answers, based on the records each one may see.Durable workflows: a task spans hours, waits for approvals or touches several systems in sequence.Operational observability: the business needs logs, alerts, cost tracking and audit trails for every answer.Any one of these points toward custom AI agents built on an API, because those give engineers full control of identity, state and monitoring.
Data questions are a special case, though. When users mostly ask about tables in a lakehouse, an assistant inside the data platform is often the better host. The Databricks Assistant is one example.
How Kanerika Helps Teams Get More From Custom GPTs
Kanerika is an OpenAI Select Partner and is Certified in the Claude Partner Network, so its teams work across both ecosystems. For GPT estates, the work follows five stages. It starts with an assessment that inventories every GPT and its exposure to the Dec 11 retirement.
Design comes next, deciding for each GPT whether a plugin, a Copilot Studio agent or a custom agent fits best. Then the build stage prepares knowledge files, rebuilds actions as MCP servers and creates the evaluation sets that prove nothing broke. Governance and enablement close the loop, so owners, review dates and user training survive the move. More on the partnership sits in Kanerika’s OpenAI service partner overview.
The proof comes from production work, none of it built as a custom GPT. Kanerika’s LLM-driven ticket response system reached 80% ticket auto-resolution after the team built a knowledge base and prepared historical tickets. A sales enablement project with Microsoft Copilot, delivered through Teams and an intranet chatbot, brought 90% fewer data query errors.
Each started, like most GPTs, with a knowledge set and a clear job, then went further because the job needed integrations, routing and measurement. Teams planning that step can explore Kanerika’s generative AI services or its agentic AI services .
Case Study
90% Fewer Data Query Errors with MS Copilot
Kanerika gave a sales team Microsoft Copilot through Teams and an intranet chatbot, with Azure SQL for real-time data, cutting data query errors by 90%.
Read the Case Study → What to Do With Your Custom GPT Before Dec 11, 2026
A custom GPT is ChatGPT configured with instructions, knowledge, capabilities and either apps or actions. You can still build or edit one where a managed workspace allows it, using the eight steps above. Personal accounts can only use and edit what they already have.
Either way, every GPT retires on Dec 11, 2026, or Feb 11, 2027 for FedRAMP and approved deferrals. So publish your final version and migrate it to a plugin. Rebuild custom actions first, because they are the one part that will not follow on its own.
Frequently Asked Questions
What is a custom GPT?
A custom GPT is a version of ChatGPT configured for one purpose. It combines written instructions, uploaded knowledge files, selected capabilities such as web search, and optional apps or actions that reach other systems. Building one does not train a new model, and each conversation with it starts fresh without memory of earlier chats.
Are custom GPTs being discontinued?
Yes. OpenAI has scheduled custom GPTs to retire on Dec 11, 2026. FedRAMP workspaces and Enterprise workspaces with an approved deferral retire later, on Feb 11, 2027. After the retirement date, GPTs and their pages become inaccessible, although your existing conversations with them remain readable. OpenAI is helping creators move them to plugins.
Can I still create a custom GPT for free or on ChatGPT Plus?
No. New GPT creation and publishing are no longer available on personal accounts, including Free, Go, Plus and Pro. You can still use existing GPTs, and edit your own if your plan allows it. Creation now depends on a Business, Enterprise or Edu workspace that still has it enabled, and OpenAI points new workflows to Plugin Creator.
What happens to my custom GPT after December 11, 2026?
On the retirement date, the GPT and its page stop working for everyone who used it. Past conversations stay accessible, but the configuration does not. To keep the assistant, migrate it to a plugin before then. The plugin keeps the instructions, knowledge files and connected apps, while custom actions have to be rebuilt separately.
How do I migrate a custom GPT to a plugin?
Publish your latest edits first, because migration copies only the most recent published version. Then open My GPTs and select Migrate to plugin. Instructions become a skill and knowledge files become reference files. The plugin starts private, so share it again, rebuild any custom actions, and compare its answers with your saved test questions.
How many files can I upload to a custom GPT?
A custom GPT accepts up to 20 knowledge files, and each file can be as large as 512 MB. Size is rarely the real limit, though. Clean, text-forward documents with clear headings give better answers than scanned pages or complex layouts, and removing outdated versions prevents the GPT from quoting two conflicting policies.
Can the creator of a custom GPT see my chats?
No. OpenAI states that GPT builders cannot view the individual conversations users have with their GPTs. Data you send through an app or action may reach that third-party service, though, and OpenAI does not audit how those services store it. On Business, Enterprise and Edu plans, data is not used for training by default.
Can I embed a custom GPT on my website?
No. Custom GPTs are designed to work inside ChatGPT and cannot be embedded in an external website or application. To put an assistant on your own site or in a product, OpenAI directs builders to its API, where the Responses API now replaces the retired Assistants API. Copilot Studio agents can also publish to websites.
What is the difference between a custom GPT and a ChatGPT Project?
A ChatGPT Project groups your own chats and files around ongoing work, so it suits a personal or small-team workstream. A custom GPT packages instructions, knowledge and capabilities into a tool that many people use the same way. Projects organize your work, while a GPT standardizes a task for others.
Does a custom GPT train a new model?
No. A custom GPT never changes the underlying model’s weights. It wraps an existing OpenAI model in standing instructions and gives it files to search, so its behavior changes without any training. If you need the model itself to learn a style or format, that requires fine-tuning through the API instead.
