If you use a personal ChatGPT account, you cannot create a new Custom GPT today. OpenAI's current guidance says creation and publishing are unavailable on Free, Go, Plus, and Pro accounts. Existing GPTs still work, while eligible Business, Enterprise, and Edu workspaces can keep building them when administrators allow it. Check your account before following any tutorial.
There is another important correction. Agent Builder is no longer the destination for new builds. OpenAI is winding it down and will remove it from the platform after November 30, 2026. The useful question is therefore broader than “How do I create a Custom GPT?” You first need to decide whether a Custom GPT, a Project, or a Workspace Agent fits the job. Agent Builder matters only if you already have a workflow to migrate.
Custom GPT, Project, or agent: choose before you build
These products solve different context and workflow problems.
| Choose | When it fits | Important boundary |
|---|---|---|
| Custom GPT | You need a reusable assistant with fixed instructions, reference files, and selected tools inside ChatGPT | Each conversation starts fresh. GPTs do not use saved memory, global custom instructions, or earlier conversations. |
| Project | You are doing ongoing work and want related chats, files, and project instructions to stay together | It is a workspace for evolving context, rather than a separately configured assistant for other people to discover and reuse. |
| Workspace Agent | An eligible organization needs a shared workflow that can use connected tools, run on a schedule, or operate through supported channels | Availability and permissions are controlled by the workspace. |
| Agent Builder | You already have a platform workflow and need to plan its exit | OpenAI says the product will no longer be available after the published shutdown date. |
The sharpest distinction is memory. Custom GPT conversations start fresh, even though the GPT's configured instructions and knowledge remain available. Projects keep chats, files, and project instructions together with built-in project memory. If you are writing a report over several sessions or accumulating research, use a Project.
A Custom GPT still suits a stable, repeatable conversation. Examples include an internal policy explainer, a tutor that follows a defined rubric, or a copy reviewer that applies the same style guide each time. Its value is the reusable configuration, not continuity between sessions.
For work that must move across systems, check Workspace Agents before building another GPT. OpenAI documents Workspace Agents as shared workflows that can connect to apps and tools, run on schedules, and be triggered through an API. If you are evaluating the older visual workflow product, read our separate guide to ChatGPT Agent Builder, then plan around its announced shutdown.
What happens to existing Custom GPTs?
They do not disappear because new personal-account creation is closed. OpenAI says existing GPTs remain usable. Whether you can edit one depends on your current subscription and workspace permissions.
Two behaviors matter when you maintain an older GPT:
- It does not remember prior GPT conversations. Put durable rules in Instructions and durable reference material in Knowledge.
- Its recommended model is not a permanent pin. OpenAI says a GPT may be moved to a similar current model when an older model is retired.
OpenAI's Agent Builder migration notice sends code-based workflows to the Agents SDK and natural-language workflows to Workspace Agents. If you need scheduled runs, shared connections, or work across tools, rebuild the workflow in the product that supports those requirements rather than assuming the old configuration will migrate.
How to create a Custom GPT if your workspace is eligible
Creation happens on the web. OpenAI's mobile apps can use GPTs but do not provide the builder. The official creation flow starts in the GPTs area of ChatGPT.
- Open Explore GPTs in the ChatGPT sidebar.
- Select Create. If the button is missing, ask your workspace administrator whether your role can build GPTs.
- Choose the conversational builder to describe the job in plain language, or open the configuration view and fill in the fields directly.
- Add a clear name, a short description, and realistic conversation starters.
- Write the Instructions, attach any Knowledge files, and enable only the capabilities the job requires.
- Use Preview with representative prompts, revise the configuration, then select Create.
- Open Share and use only the access levels your workspace makes available.
The button sequence is the easy part. Most of the quality comes from separating four configuration layers correctly.
Instructions, Knowledge, Capabilities, and Actions
Instructions define behavior
Instructions tell the GPT what job to do, how to make decisions, how to format the answer, and when to stop or ask for help. OpenAI recommends explicit steps, concrete directions, and examples when the GPT must apply a definition or classification.
Start with an operational template:
You are a customer support assistant for [company].
Use the uploaded knowledge before answering.
If the source material does not support an answer, say what is missing and route the user to [owner or channel].
Never invent pricing, legal, medical, or security information.
For each answer:
1. Give the direct answer.
2. Name the source document used.
3. State any action the user needs to take.
Replace the placeholders with a real owner and escalation path. A refusal without a next step merely transfers the problem back to the user.
Knowledge supplies reference material
Knowledge is for source material such as policies, manuals, product documentation, and approved examples. Rules about tone or workflow belong in Instructions. OpenAI also recommends clear, text-forward files and testing retrieval in Preview.
Before uploading a document, check its owner, revision date, and whether an authoritative version already exists. Remove superseded drafts and duplicated policies. If the source changes, the GPT's uploaded copy does not become current on its own.
Capabilities add built-in tools
Capabilities can include web search, image generation, Canvas, and Code Interpreter & Data Analysis, depending on the account, workspace, and region. Enable a capability because a test case requires it. A policy assistant grounded in approved internal files may not need web search; a spreadsheet reviewer may need data analysis.
Actions connect external APIs
Actions let a GPT call external APIs that you define. Apps use services a user has connected. A GPT can use apps or actions, but not both at the same time, so decide which integration model you need before configuring the assistant.
A Custom GPT also cannot be embedded as a chat widget on your site. OpenAI directs website and product integrations to its API. That boundary was missing from the old article and changes the deployment decision.
Test the job, the boundaries, and the sources
Preview should cover the work you expect and the failures you cannot accept. Build a small evaluation set from real task shapes, with sensitive details removed.
Include:
- A normal request with a clearly supported answer
- An ambiguous request that should trigger a clarifying question
- A request outside the GPT's assigned scope
- A conflict between its instructions and an uploaded document
- A request for a claim the knowledge files do not support
For each response, check the decision, the cited source, the output format, and the escalation behavior. If the same failure repeats, repair the relevant layer. Change Instructions for a behavior problem, Knowledge for missing or obsolete source material, and capabilities only when the GPT lacks a necessary tool.
OpenAI's editor provides Preview and version history, including the ability to restore an earlier version. Use the version description to record what changed and which evaluation case justified it. If you restore a version that used actions, OpenAI warns that its authentication may need to be configured again.
Share only after you know the audience
Sharing options vary by account and workspace policy. An eligible managed workspace may allow access for selected people, link holders, the workspace, or the GPT Store. Administrators can restrict those choices. Review the current sharing and publishing rules before distribution.
Use the narrowest audience that can complete the task. A GPT with internal policies or connected services belongs inside the governed workspace. Public publishing introduces additional policy and integration requirements. A public GPT that uses actions needs a valid privacy policy URL for each public action.
The practical answer in 2026
Create a Custom GPT only if you are in an eligible managed workspace and need a reusable assistant inside ChatGPT whose conversations can start fresh. Use a Project when the value lives in accumulated chats and files. Use a Workspace Agent when an eligible team needs a shared workflow with tools, schedules, or supported external channels.
Do not begin new work in Agent Builder. Migrate an existing workflow before OpenAI's announced cutoff, choosing the Agents SDK for a code-based implementation or a Workspace Agent for a supported natural-language workflow. That decision will outlast any step-by-step tour of a button that may no longer be there.