Custom GPT
Last reviewed
Sources
No citations yet
Review status
Needs citations
Revision
v6 ยท 4,322 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
No citations yet
Review status
Needs citations
Revision
v6 ยท 4,322 words
Add missing citations, update stale details, or suggest a clearer explanation.
A custom GPT (branded by OpenAI simply as a "GPT") is a tailored, no-code version of ChatGPT that a person configures for a specific purpose by giving it custom instructions, knowledge files, and selected tools, then keeps private, shares by link, or publishes for others to use. It is not a separate AI model: a custom GPT is a saved configuration that runs on top of an existing base model in OpenAI's GPT family inside the ChatGPT product. OpenAI announced GPTs on November 6, 2023 at its first developer conference, and within two months users had created more than 3 million of them.[1][2] The headline pitch is that building one requires no programming: "Anyone can easily build their own GPT, no coding is required."[1]
A custom GPT bundles a name, a description, a set of natural-language instructions, optional conversation starters, optional uploaded knowledge files, a choice of built-in capabilities (web search, image generation, a Python sandbox, and so on), and optional "actions" that let it call external APIs. OpenAI opened the public GPT Store for sharing them on January 10, 2024, by which point the more than 3 million figure had already been reached.[1][2] Because a custom GPT is a configuration layered on a model in the GPT-4 family rather than a model itself, the distinction between "a custom GPT" (something a user built) and "GPT-4" or "GPT-4o" (the underlying large language model) matters and is easy to blur.
This article covers the user-built assistant. The public marketplace has its own page at GPT Store, and the older extension mechanism that custom GPTs partly replaced is covered at ChatGPT plugins.
OpenAI's own one-line definition, from its Help Center, is that "GPTs (also called custom GPTs) are versions of ChatGPT configured for a specific purpose," combining "specific instructions, knowledge, and selected capabilities to create a more tailored experience in ChatGPT."[3] In practice a custom GPT is a stored prompt-and-tools profile. When a user opens the GPT and sends a message, ChatGPT runs the conversation using the builder's configuration as a hidden system prompt, switches on only the tools the builder enabled, and pulls from the builder's uploaded files when relevant.
The idea OpenAI pitched at launch is that a genuinely useful assistant usually needs three things the plain chatbot cannot supply on its own: domain instructions, private knowledge, and the ability to take action in another system. A custom GPT lets someone who cannot program combine those pieces and hand the result to colleagues, customers, or strangers. OpenAI framed GPTs as a successor to Custom Instructions, the preference settings it had shipped in July 2023, noting that "many power users maintain a list of carefully crafted prompts and instruction sets, manually copying them into ChatGPT. GPTs now do all of that for you."[1]
It is worth being precise about what a custom GPT is not. It does not fine-tune or retrain the base model; the model weights are untouched. It does not run outside ChatGPT; OpenAI is explicit that GPTs "are not a way to embed ChatGPT in an external website or application," and that anyone wanting an assistant inside their own product should use the API instead.[3] And it does not carry your personal context: GPTs "do not use saved memory, custom instructions, or previous conversations," so each conversation with a GPT starts fresh.[3]
A custom GPT is a configuration, not a model. GPT-4, GPT-4o, and GPT-5 are the underlying large language models that ChatGPT runs on; a custom GPT is a saved profile (instructions, knowledge files, tool selections, and optional actions) that sits on top of whichever base model ChatGPT offers at the time. Building one does not change the model's weights, and OpenAI is explicit that GPTs "do not use saved memory, custom instructions, or previous conversations."[3] The practical difference is that two custom GPTs running on the same base model can behave very differently because of their instructions, knowledge, and enabled tools, while neither is a distinct model in its own right. When a base model a GPT relied on is retired, ChatGPT automatically substitutes a similar current model, which is why builders set a "recommended model" rather than hard-code one.[3][7]
OpenAI introduced GPTs during the opening keynote of DevDay, its first developer conference, held in San Francisco on November 6, 2023. The launch post described them as "custom versions of ChatGPT that you can create for a specific purpose," able to "help you learn the rules to any board game, help teach your kids math, or design stickers."[1] At launch, example GPTs were available to ChatGPT Plus and Enterprise subscribers, including ones from Canva and Zapier, and OpenAI said it planned "to offer GPTs to more users soon."[1]
The same keynote introduced the model and developer tooling that custom GPTs initially ran on top of: GPT-4 Turbo, a model with a 128,000-token context window, and the Assistants API. OpenAI noted that the Assistants API "is built on the same capabilities that enable our new GPTs product: custom instructions and tools such as Code interpreter, Retrieval, and function calling."[4] In other words, a custom GPT and an API-built assistant are two front ends over the same underlying machinery; the GPT is the no-code, in-ChatGPT version, and the Assistants API is the developer version meant to live in an external product.
OpenAI opened the GPT Store on January 10, 2024, "two months since we announced GPTs," at which point "users have already created over 3 million custom versions of ChatGPT."[2] The store rolled out to ChatGPT Plus, Team, and Enterprise users. It launched a few weeks later than the "later this month" target Sam Altman had floated on the DevDay stage, a slip widely attributed to the brief November 2023 board crisis that removed and then reinstated Altman as CEO.[1][2]
The launch was paired with a new paid tier, ChatGPT Team, announced the same day, which gave smaller organizations a private section of the store for GPTs published only to their workspace.[2]
Toward the end of January 2024, OpenAI rolled out "GPT mentions," which let a paying user pull a specific GPT into an ordinary conversation by typing @ followed by the GPT's name. The selected GPT then answers inside the same thread, and, as OpenAI's documentation puts it, "the conversation keeps its current context."[3][5] The practical effect is that you can chain specialists: ask a research GPT for sources, then @ a writing GPT to turn them into a draft, without losing the thread. TechCrunch reported the feature reaching ChatGPT Plus users around January 30, 2024.[5]
When OpenAI introduced GPT-4o on May 13, 2024, it also began bringing previously paid features to free accounts, including the ability to browse the GPT Store and use GPTs built by others.[6] The split that resulted still holds: anyone signed in to ChatGPT can use a custom GPT, but building or editing one requires a paid subscription.[3] Free users also hit a message cap, after which ChatGPT falls back to a smaller model.[6]
Custom GPTs follow whichever base models ChatGPT offers, and that lineup shifts. As of February 13, 2026, OpenAI had retired GPT-4o, GPT-4.1, GPT-4.1 mini, OpenAI o4-mini, and the Instant and Thinking variants of GPT-5 from ChatGPT, while keeping the underlying API access unchanged. As a transition measure, Business, Enterprise, and Edu customers were told they would retain access to GPT-4o specifically inside custom GPTs until April 3, 2026, after which it would be fully retired across all plans.[7] When a model a GPT relied on disappears, ChatGPT automatically switches that GPT to a similar current model, which is why builders are encouraged to set a "recommended model" rather than hard-code expectations about a particular one.[3][7]
Building and editing GPTs happens only on the web version of ChatGPT; the mobile apps can use GPTs but cannot build them. The editor opens from "Explore GPTs" in the sidebar or at chatgpt.com/gpts, and it offers two ways to work.[7]
| Mode | What it does |
|---|---|
| Conversational builder | ChatGPT interviews the builder, asking what the GPT should do, and drafts a name, an icon, conversation starters, and a first set of instructions automatically. |
| Configuration view | A direct form where the builder edits each field by hand: name, description, instructions, conversation starters, knowledge, capabilities, and actions. |
In practice many serious builders draft in the conversational mode and then switch to the configuration view to rewrite the instructions, because the auto-generated prompt tends to be generic.
Instructions are the free-form natural-language prompt that defines "how your GPT behaves: what it should do, how it should respond, and what it should avoid," and they apply to every conversation.[7] OpenAI's own guidance is to prefer "positive, concrete instructions" over long lists of prohibitions, to use explicit step structure for multi-step workflows, and to include short examples of acceptable and unacceptable output when the GPT must apply specific rules or classifications.[7]
A builder can attach up to 20 files to a single GPT, each up to 512 MB, covering "most common document, spreadsheet, image, text, and code file types."[7] When a user asks something, ChatGPT searches those files using retrieval and feeds the relevant chunks back into the prompt. Some file types only work when the Code Interpreter and Data Analysis capability is turned on.[7]
Retrieval is not flawless. For a large document, asking about a particular page or table can still produce a wrong or made-up answer, which is why builders often split their material into smaller, text-forward files before uploading. OpenAI's blunt advice is to "prefer clear, text-forward files when possible," since complex layouts make uploaded content harder to use.[7] Knowledge files are meant to stay private to the builder; a user can see excerpts the GPT chooses to quote, but not the raw files. That privacy is not absolute, which the security section below covers.
Capabilities are the built-in tools a builder can switch on or off. Locking a GPT to a subset is part of the point: a writing assistant can disable code execution to keep the experience focused. The current set in the editor is:[7]
| Capability | What it adds |
|---|---|
| Web search | Retrieves up-to-date information from the live web. |
| Image generation | Creates images from text prompts (historically via DALL-E). |
| Canvas | A side panel for drafting and editing longer or structured content. |
| Code Interpreter and Data Analysis | Runs Python in a sandbox for calculations, data work, and charts. |
| Apps | Lets the GPT use external tools and services the user has connected. |
These are the same tools available in regular ChatGPT, just scoped to whatever the builder selected. Availability varies by account, workspace, and region.[7]
Actions are how a custom GPT reaches the rest of the internet. The builder supplies an OpenAPI specification (in JSON or YAML) describing a third-party API plus an authentication setup, and at runtime ChatGPT uses function calling to decide when a call is needed, fills in the request, sends it, and reads the response back into the conversation.[8] OpenAI summarizes the schema as telling ChatGPT "which server to call, what endpoints are available, what parameters they accept, and how each action is identified."[8]
Authentication can be one of three kinds:[8]
| Auth type | Notes |
|---|---|
| None | No credentials; suitable for open, public endpoints. |
| API key | A secret key sent as Basic, Bearer, or a custom header; meant for server-to-server access. |
| OAuth | User sign-in for account-scoped access, requiring a client ID and secret, authorization and token URLs, and scopes. |
The editor ships starter schemas (a Weather example in JSON, a Pet Store example in YAML, and a blank template), validates the schema and shows the detected operations, and provides a callback URL for completing an OAuth flow.[8] Two limits are worth knowing: a GPT can use either Apps or Actions but not both at once, and actions are not available in ChatGPT's Pro mode, so the model picker hides Pro models when a GPT has custom actions.[7][8]
Actions are the direct successor to the older ChatGPT plugins system. OpenAI said the design "builds upon insights from our plugins beta," and that "migrating from the plugins beta is easy with the ability to use your existing plugin manifest to define actions for your GPT."[1] The architectural difference is about ownership: a plugin was a global add-on that a user toggled per conversation, whereas an action lives inside one specific GPT, which decides on its own when to call out. The plugin system itself was wound down in 2024 (new conversations with plugins stopped on March 19, and all plugin chats ended on April 9), with OpenAI pointing former plugin users at the GPT Store and former plugin builders at actions.[9]
Conversation starters are up to a handful of short example prompts shown as buttons when a user opens the GPT. They are not part of the instructions; they are user-facing hints that help people figure out what the GPT is for, and a good set is often the difference between a GPT that feels usable and one that feels like an empty text box.[7]
Every GPT also has a profile (name, description, icon, builder name) and three visibility settings: private to the builder, anyone with the link (unlisted), or public in the GPT Store. The editor keeps a version history so a builder can review and restore earlier versions, although restoring a version that used actions may require reconfiguring its authentication.[7]
| Element | Required? | Limits and notes |
|---|---|---|
| Name and description | Yes | User-facing; shown in search, shared links, and store listings. |
| Instructions | Effectively yes | Free-form system prompt applied to every conversation. |
| Conversation starters | Optional | A few example prompts displayed as buttons. |
| Knowledge files | Optional | Up to 20 files, 512 MB each; retrieval-based. |
| Capabilities | Optional | Web search, image generation, Canvas, Code Interpreter and Data Analysis, Apps. |
| Actions | Optional | OpenAPI schema (JSON or YAML); None / API key / OAuth; not available in Pro mode; mutually exclusive with Apps. |
| Recommended model | Optional | Suggests a base model; ChatGPT substitutes a similar one if it is unavailable. |
| Version history | Built in | Review and restore earlier versions. |
Who can do what with a custom GPT depends on the plan. Using a GPT that someone has shared is free for any signed-in ChatGPT user, but building or editing one requires a paid subscription.[3][6]
| Action | Free | Plus / Go / Pro | Team / Enterprise / Edu |
|---|---|---|---|
| Use a public or shared GPT | Yes (signed in) | Yes | Yes |
Invoke a GPT with @ in a chat | Limited rollout | Yes | Yes |
| Build and edit GPTs | No | Yes | Yes (subject to workspace permissions) |
| Publish publicly in the GPT Store | No | Yes (verified builder) | Workspace-private publishing |
| Conversations used to train models | Possible unless opted out | Possible unless opted out | No, by default |
Using a GPT requires being signed in; public GPT pages may be visible to logged-out visitors, but ChatGPT prompts for sign-in before chatting.[3] Building requires a paid subscription, and in managed Enterprise or Edu workspaces an admin can further restrict who may create or edit GPTs and which sharing methods are allowed.[3] If a builder downgrades or cancels a qualifying plan, they keep the ability to use the GPTs they already made but lose the ability to edit them or create new ones.[3]
The GPT Store is the public directory for custom GPTs, reachable from inside ChatGPT. It sorts public GPTs into categories (OpenAI's launch list was DALL-E, writing, research, programming, education, and lifestyle), runs a community leaderboard ranked by usage, and features a rotating set of editor-picked GPTs.[2] The launch lineup OpenAI highlighted on January 10, 2024 illustrates the intended range:
| Featured GPT | Builder | Purpose |
|---|---|---|
| AllTrails | AllTrails | Personalized trail recommendations |
| Consensus | Consensus | Search and synthesize results from 200 million academic papers |
| Code Tutor | Khan Academy | Help learners build coding skills |
| Canva | Canva | Design presentations and social posts |
| Books | OpenAI | Find your next read |
| CK-12 Flexi | CK-12 Foundation | An AI tutor for math and science |
To list a GPT publicly, a builder must save it for "Everyone," verify a Builder Profile (a real name or a verified website), and pass review against OpenAI's usage policies and brand guidelines. That review "includes both human and automated review," and users can report GPTs that break the rules.[2]
On monetization, OpenAI said at the store launch that "in Q1 we will launch a GPT builder revenue program," and that "as a first step, US builders will be paid based on user engagement with their GPTs," with payment criteria to follow.[2] The program entered a pilot in March 2024, limited to a small group of US-based builders and tied to how much paying users engaged with their GPTs rather than to per-call usage or subscription fees.[2][10] The exact payout formula has never been published in full, and the program has stayed US-only: builders elsewhere can publish and rank in the store but cannot collect payouts.[10]
OpenAI's privacy framing for GPTs rests on a few specific guarantees and one recurring caveat.[1][3]
The guarantees: a builder cannot read the individual conversations users have with their GPT.[3] Whether those conversations are used to improve OpenAI's models depends on the plan, not the GPT: on Business, Enterprise, and Edu plans, data is not used for training by default, while on consumer plans (Free, Plus, Go, Pro) it may be used unless the user opts out in data controls.[3] And when a user is signed in to a managed workspace, OpenAI does not use conversations with GPTs to improve its models at all.[1][2]
The caveat is about reaching outside ChatGPT. When a GPT calls a third-party API through an action, or uses a connected app, "relevant parts of your input may be sent to the third-party service," and "OpenAI does not audit or control how those services use or store your data."[3] ChatGPT may ask the user to approve a request before it runs, and every public GPT that uses actions must publish a privacy policy URL, but the practical guidance is plain: only use GPTs whose APIs and apps you trust.[3][8]
The most-discussed weakness of custom GPTs is that their two "private" pieces, the instructions and the knowledge files, are not reliably private against a determined user. Through prompt-injection and "jailbreak" prompts, people have repeatedly coaxed widely shared GPTs into printing their own system prompt or leaking the contents of uploaded files. In 2024, researchers at Palo Alto Networks and several academic groups documented instruction-leak and file-extraction attacks against popular GPTs and advised organizations to treat anything uploaded to a public GPT as eventually disclosable.[11] The takeaway builders are warned about is simple: do not put secrets in instructions, and do not upload truly sensitive files to a GPT you intend to share.
Retrieval quality is the other built-in limit. Because knowledge files are searched rather than loaded whole, a GPT can miss the one passage that answers a question, especially in long or densely formatted documents. And like the base chatbot, a custom GPT can still hallucinate, follow injected instructions from a retrieved document, or call an action with the wrong arguments.
The store grew faster than OpenAI's ability to police it, and the gap was visible early. A March 2024 TechCrunch investigation found the GPT Store "filling up with spam" and with GPTs that "very transparently" broke the rules, including ones that simulated conversations with real public figures (searches for names like Elon Musk, Donald Trump, Barack Obama, and Joe Rogan turned up dozens), despite OpenAI's policy forbidding GPTs that impersonate people or organizations without "consent or legal right."[12] OpenAI's stated enforcement model combines automated systems, human review, and user reports, with consequences ranging from warnings and sharing restrictions to removal from the store and loss of monetization eligibility.[12] Critics note the model is largely reactive: offending GPTs are removed in batches after they are reported, by which point users have often already used them.
A separate, more mundane complaint concerns discovery. Rankings are opaque, the leaderboard tends to favor GPTs that arrived early, and near-duplicate clones of popular GPTs (sometimes with almost identical names) regularly appear and confuse users.
Beyond consumer novelties, the most durable adoption has been inside companies, where internal-only GPTs sidestep the public store entirely. At the DevDay launch OpenAI cited early enterprise users such as Amgen, Bain, and Square building internal GPTs for marketing copy, customer support, and engineer onboarding.[1] The most cited example is the pharmaceutical company Moderna, which OpenAI's own case study says deployed more than 750 GPTs across functions including legal, research, manufacturing, and commercial, with one GPT (Dose ID) used to help evaluate vaccine dosing; Moderna reported that roughly 40 percent of its weekly ChatGPT Enterprise users interacted with GPTs.[13]
Custom GPTs are the most visible example of a broader category that vendors call configurable assistants. The closest analogues are Anthropic's Claude Projects and Google's Gemini Gems. The shape of the trade-off is reach versus context: the GPT Store gives builders a real audience and the prospect of payment, while Claude Projects and Gemini Gems stay private to a user or team. On the technical side, custom GPTs rely on retrieval over uploaded files, which scales to larger collections but can miss a specific passage, whereas Claude Projects tend to load files directly into a long context window, which is usually more reliable for questions about a particular page.
OpenAI has also been folding the simpler use cases back into ChatGPT itself, through Custom Instructions, memory, and Projects (a folder-style workspace that bundles files, instructions, and chats without publishing a public GPT). These features overlap with custom GPTs and increasingly absorb the lightweight cases that once justified spinning up a private GPT.
Custom GPTs were OpenAI's first mass-market attempt to turn ChatGPT from a single chatbot into a platform that anyone could extend and that anyone could distribute. The launch post was candid that this was a deliberate step toward AI "agents," systems that "take on real tasks in the real world," and that OpenAI wanted to "move incrementally towards this future."[1] Whether the GPT Store became the App Store of AI is debatable; usage concentrated among a relatively small set of GPTs, monetization stalled, and moderation lagged. The more lasting contribution may be the actions pattern, which normalized describing a tool to a model with an OpenAPI schema and letting the model decide when to call it. That same arc, away from vendor-specific manifests and toward a model-agnostic protocol for tool use, continued with Anthropic's Model Context Protocol, introduced in November 2024 and later adopted across OpenAI's products, which lets ChatGPT talk to external tools without wrapping each one in a bespoke GPT.[14]