You can access ChatGPT's assistant-building space in two places, and both require a paid plan. The first is the GPT editor at chatgpt.com/create, where you assemble a custom GPT from instructions, knowledge files, and actions. The second is Projects, a workspace you create from the left sidebar. Free accounts see neither builder; Plus, Pro, Business, Enterprise, and Edu accounts see both. The boundary is the subscription, not a hidden menu.

The Two Building Spaces ChatGPT Actually Offers

ChatGPT splits custom assistant creation across two products that get confused constantly. A custom GPT is a standalone assistant you publish, share by link, or keep private, and it carries its own name, icon, instructions, conversation starters, and knowledge files. A Project is a container that groups chats around one body of work, with project-only memory and optional uploaded files.

The entry points differ. The GPT builder lives at chatgpt.com/create and inside the left sidebar under "GPTs," where a "Create" button opens the editor. Projects appear as a "New project" option in the same sidebar once your account is entitled. Remember one distinction: GPTs change how the model answers everyone who uses them, while Projects change what one workspace remembers for you.

ChatGPT also hosts a third space most people never open: the Assistants playground on the developer platform, at platform.openai.com. That is the API product, not the ChatGPT app, and it bills per token rather than through your subscription. If you landed here expecting to drag sliders inside a chat window, the playground will feel like a different country with its own currency.

Access depends on plan, region, and platform. OpenAI's help center publishes a plan comparison page that lists custom GPT creation and Projects as paid features, and that page is the authority to check, because entitlements shift between plans and countries faster than any article survives.

Who Gets In: Plan and Account Requirements Behind the Builder

The gate is the subscription. Free accounts can use GPTs other people published, but the "Create" button is absent, and OpenAI's help center documents custom GPT creation as a Plus-and-above feature. Projects arrived later and rolled out across paid tiers; the exact tier list sits on the same help center page. Check it before you upgrade for this one feature.

Business, Enterprise, and Edu workspaces add a second gate above the subscription: the admin. Workspace owners can restrict GPT creation, and members of an Enterprise workspace may find the option missing until an admin enables it in workspace settings. Admins should confirm the toggle in the admin console before assuming a member has the same access they do.

Under-eighteens face an age gate. OpenAI's terms require users to be at least 13, and its teen safety policies restrict account activity for minors further, which can remove builder access even on a paid account. A parent-managed teen account is not the same as an adult Plus subscription.

One platform caveat: the ChatGPT desktop apps for macOS and Windows mirror the web interface closely, but the GPT editor is defined by the web experience at chatgpt.com. If a menu item is missing on mobile, open chatgpt.com in a browser before concluding your account lacks the feature. Screen size hides buttons; entitlements hide sections.

Getting Into the GPT Editor From a Cold Start

Work through the sequence in order; each step assumes the previous one worked.

1: Sign in at chatgpt.com in a desktop browser with a paid account. The left sidebar loads with Chats, Projects, and GPTs among its entries. 2: Click "GPTs" in the sidebar. If you do not see it, your plan lacks creation rights, not just visibility. 3: Click "Create" in the upper-right of the My GPTs page. The editor opens at chatgpt.com/create with two panes: a chat-style configuration panel and a live preview. 4: If the editor offers a "Create" tab and a "Configure" tab, click "Configure." This exposes the manual fields the guided chat hides. 5: Fill the Name and Description fields. Both are required before the Save button activates in most builds. 6: Paste your instructions into the Instructions box (the field's limit is 8,000 characters; anything longer needs an uploaded file instead). 7: Leave "Web Search" and "Canvas" unchecked unless the task needs them β€” each capability you enable changes how the GPT responds and expands its attack surface. 8: Click "Create" in the upper-right, choose "Only me," "Anyone with a link," or "GPT Store," then click "Save."

Failure handling: if the editor loads but Save stays greyed out, a required field is empty β€” usually Description. If clicking "Create" sends you to an upsell page, you are signed into a free workspace inside your account; switch workspaces from the account menu and retry. If the editor throws "Something went wrong," reload chatgpt.com/create once, and if the error repeats, sign out and back in before opening a support ticket.

What Each Configure Field Controls, and Where Builds Go Wrong

The Instructions box is where the assistant's behavior actually lives, and it accepts up to 8,000 characters in the current editor. Everything after that is plumbing: Conversation starters (up to four prompts shown to users), Knowledge (up to 20 files in the editor's uploader for Plus accounts), and Capabilities (web search, image generation, and code execution each toggle separately). Under "Actions," you can attach a public API schema; an Action without a valid authentication method fails silently at runtime, not at save time.

The most common build failure is an instruction file that contradicts itself. Instructions carry the highest priority of anything in the GPT, so stating "never mention pricing" in line one and "quote our prices" in line nine produces inconsistent answers rather than an error. Keep the rules list short and non-overlapping.

Knowledge files get retrieved, not memorized. If your uploaded PDF is scanned images without a text layer, the GPT cannot read it, and users will get confident-sounding fabrications instead of your document. Convert scans through OCR before upload.

Conversation starters are easy to neglect and easy to misuse. A starter phrased as a question ("What's your return policy?") tempts users to test the GPT with unrelated queries against the same starter. Phrase starters as tasks ("Draft a return-policy reply for this customer message:").

Version state matters here. The GPT editor's exact field layout has changed across ChatGPT releases, and OpenAI's help center article on creating a GPT reflects the current fields; build from that article's field list when your screen disagrees with any guide older than a year.

Building a Project Workspace and What It Remembers

A Project opens from the left sidebar with "New project," and the workspace asks for a name before you type anything else. Inside, the project groups its own chat threads, keeps project-only memory separate from your general ChatGPT memory, and accepts uploaded reference files that stay scoped to that project β€” your other chats cannot see them.

Project memory is the feature that makes Projects useful for repeated work. A chat inside the project recalls earlier decisions and drafts from the same project, but chats outside it cannot, and deleting the project removes that scoped memory. If you need fidelity to two sets of rules β€” say, a client's brand voice and your own β€” keep them in separate projects with separate memory.

Determining which tool to use takes one question: does the assistant answer everyone the same way, or does it recall your specific work? A custom GPT answers everyone the same way; a Project recalls yours. Most professional workflows use both β€” a GPT for the reusable method, a Project for the specific engagement.

Projects also accept file uploads for reference context, though the uploader's size and count limits are set by ChatGPT plan and change; check the file-attachment help article for the current numbers before planning around a large corpus. For heavy document work, upload fewer, better-normalized files and confirm the project can quote from them correctly before inviting teammates.

The API Playground: The Other Space, Its Billing, and Its Ceiling

If you want programmatic control β€” a fixed system prompt, tool definitions, and an assistant you call from your own app β€” the space is the Assistants product on OpenAI's developer platform, reached through platform.openai.com. There you create an assistant object, choose a model, attach files, and define function tools. It is not the ChatGPT app; it does not appear in your sidebar or in the GPT Store.

Billing is the difference that trips people up. The playground bills per token against API credits, separate from a ChatGPT subscription; a Plus subscription does not include API usage, and API credits do not include ChatGPT Plus. Teams sometimes budget for one and discover they need the other mid-project.

Attribution matters here too. The Assistants API's capabilities, model options, and pricing live in OpenAI's API documentation and pricing page, and those are the pages that govern the product's behavior β€” not this article, and not third-party tutorials from older model generations.

The practical rule: pick the GPT editor when non-technical colleagues must use the assistant inside ChatGPT, and pick the API when software must call the assistant without a human in the chat window. Building a GPT first and porting its instructions to an API assistant later is a legitimate path, and the instructions transfer nearly unchanged.

Side-by-Side: Custom GPT, Project, and API Assistant Compared

All three live under the same OpenAI account and the same login, but they diverge sharply on who can use the result and what it costs to run.

Dimension Custom GPT Project API Assistant
Where you build it chatgpt.com/create "New project" in the left sidebar platform.openai.com playground
Minimum plan Paid ChatGPT plan Paid ChatGPT plan API account with credits
Who can use the finished thing Anyone you share it with You and workspace members Any software holding your API key
Persistent memory Instructions persist; chats do not Project-only memory persists Depends on the app you build
Knowledge files Up to 20 files in the editor's uploader Plan-dependent file uploads Files attached via the API
Publishing GPT Store, link, or private Not publishable Not publishable
Billing basis Included in ChatGPT plan Included in ChatGPT plan Per token

Read the table as three different products wearing one logo. Sharing is where the split is widest: a custom GPT can be published to the GPT Store for anyone on ChatGPT, a Project can never leave your workspace, and an API assistant can be embedded in a website without any ChatGPT interface involved.

For a small team, the cheapest configuration is usually a Business workspace with shared GPTs and per-client Projects, and API credits reserved for the one automation a human shouldn't babysit. For a solo consultant, a Plus plan with private GPTs covers both needs at the plan price.

Consider the shared-workspace bottleneck. Members of a Business workspace can build GPTs only if their admin permits it; a team that centralizes builders in an admin role will find Projects sufficient while a team that wants everyone building will need the toggle open, with review rules in place for what gets published to the Store.

If a recommended path fails, downgrade gracefully: a solo builder who cannot justify Plus can still draft instructions in a plain chat, then paste them into the GPT editor the day they upgrade. And anyone who needs the builder only occasionally should check the plan comparison page first β€” access tiers move, and a plan purchased for this feature deserves confirmation on the week you pay.

Seven Mistakes That Lock You Out or Break the Build

The first mistake is searching the sidebar for a "Create an AI assistant" label. No such label exists. The GPT builder hides behind "GPTs," and Projects hide behind "New project," so a user scanning for the literal phrase concludes the feature is gone.

The second mistake is building on a free account and expecting to save. The editor may load partially for logged-out or free sessions before the save step blocks with an upsell. Sign in with the paid account first.

The third mistake is assuming the mobile app exposes the full builder. The web interface at chatgpt.com is the reference experience; the ChatGPT apps vary by platform and version, and missing menus there are not missing features.

The fourth mistake is uploading scanned PDFs as knowledge. Retrieval reads text layers; image-only files turn into silent fabrication. Run OCR first.

The fifth mistake is setting an Action without authentication. An Action that calls an API expecting a key will fail at runtime after a clean save, and the failure looks like a broken GPT rather than a config error.

The sixth mistake is enabling every capability. Web browsing, code execution, and image generation each broaden what the assistant can do and what can go wrong; enable what the task needs and nothing else.

The seventh mistake is treating the API playground as a free version of the builder. It bills against credits separately from your subscription, and an idle experiment there can still cost money if you wire it into an app.

Sharing, Publishing, and the Privacy Line in the GPT Store

Saving a GPT offers three visitors levels: "Only me," "Anyone with a link," and "GPT Store," where anyone on ChatGPT can find it. The link option is the practical middle path for client-facing assistants, because the URL works for signed-in ChatGPT users without your assistant appearing in public search.

Privacy boundaries deserve naming. Instructions, knowledge files, and Actions travel with a shared GPT, and users who receive a link can prompt against whatever you uploaded β€” so never put credentials, internal pricing, or client data in a GPT you share. Keep sensitive material in a Project instead, where it stays inside your workspace.

The GPT Store adds discovery but also review and rights considerations. Publishing makes your assistant visible to users beyond your control, and OpenAI's usage policies govern what a published GPT may do; a GPT that impersonates a brand or gives regulated advice risks removal. Read the usage policies before choosing the Store option.

Sharing has a maintenance cost. A published GPT against an API you control can break when the API changes, and the failure surfaces to every user with your assistant's name on it. Treat each shared GPT as a product with an owner, a changelog note, and a plan for deprecation.

FAQ

Does ChatGPT have a hidden space for building assistants that free users can open?

No. The builder is not hidden β€” it is gated to paid plans, and OpenAI's plan comparison page lists custom GPT creation as paid-tier. Free accounts can use other people's GPTs but never open the editor itself.

Can I create a custom GPT on my phone?

The ChatGPT mobile apps expose GPT discovery and use, but the configuration editor with instructions, knowledge, and Actions is defined by the web app at chatgpt.com. Open a desktop browser for the full builder before concluding the feature is missing.

What is the difference between a custom GPT and a Project in terms of who can use it?

A custom GPT can be shared through a link or published to the GPT Store for other ChatGPT users; a Project stays private to your account or workspace and can never be published. Rule of thumb: sharing goes to GPTs, personal continuity goes to Projects.

Why is the Save button greyed out in the GPT editor?

Required fields are empty β€” usually the Description. Fill Name and Description, confirm the Instructions box has content, and the Create button in the upper-right activates; if it still doesn't, you may be signed into a free workspace inside a paid account.

Do I need a ChatGPT Plus subscription to use the API playground?

No. The playground at platform.openai.com bills per token against API credits, which are separate from ChatGPT subscriptions. A Plus plan gives you the GPT editor and Projects, not API usage.

What happens to a published GPT if I stop paying for ChatGPT?

Access to the builder ends, and your published GPTs stop working. Downstream users lose the assistant without warning; if a team depends on one, keep the building account funded or migrate the instructions to an API assistant you control.

Can I build an assistant that my own website calls without users visiting ChatGPT?

Yes, but through the API product, not a custom GPT. You create an assistant on the developer platform, embed the API calls in your site, and protect the key server-side; the Assistants API documentation describes the objects and endpoints involved.

Does creating a GPT require an age-verified account?

Yes. OpenAI's terms require users 13 and older, and teen safety restrictions on minor accounts can remove builder access. A parent-managed teen account does not inherit the building rights of an adult Plus subscription.