Custom GPTs
Last reviewed
May 11, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v5 ยท 3,007 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 11, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v5 ยท 3,007 words
Add missing citations, update stale details, or suggest a clearer explanation.
Custom GPTs (also called GPTs) are tailored versions of ChatGPT that anyone can build for a specific purpose without writing code. A Custom GPT bundles a name, description, a system prompt, conversation starters, optional knowledge files, and optional Actions (third-party API calls) into a single shareable assistant. OpenAI announced GPTs at its first DevDay conference on November 6, 2023, and launched the public GPT Store on January 10, 2024.
See also: ChatGPT and GPT Store
A Custom GPT is a saved configuration that runs on top of a base GPT-4 class model inside the ChatGPT product. The creator (called a "builder") writes the instructions, picks which built-in tools to enable, uploads any reference files, and optionally wires up an external API. End users open the GPT like any other chat, and ChatGPT runs the conversation using the builder's configuration as a hidden system prompt.
The key idea OpenAI pitched at launch is that a useful assistant often needs three things the base ChatGPT cannot supply on its own: domain instructions, private knowledge, and the ability to take action in another system. Custom GPTs let a non-programmer combine those pieces and hand the result to colleagues or strangers.
As of January 2024, OpenAI said users had created more than 3 million GPTs in the two months since launch. A much smaller subset, roughly 150,000 by third-party counts, is published publicly in the GPT Store. The rest are private or shared only by link.
OpenAI introduced GPTs at its first DevDay conference in San Francisco on November 6, 2023. CEO Sam Altman demonstrated the GPT Builder on stage by creating a "Startup Mentor" GPT in real time, configuring its persona through a chat dialogue rather than a form. The same keynote introduced GPT-4 Turbo, the Assistants API, and the DALL-E 3 API.
At launch, GPTs were available only to ChatGPT Plus and ChatGPT Enterprise subscribers. OpenAI also previewed a public marketplace, the GPT Store, and said builders would eventually be paid based on usage.
The GPT Store went live on January 10, 2024, two months after the original announcement. The launch was delayed from the original "later in November" timing, in part because of the brief board crisis that removed and reinstated Altman as CEO in late November 2023. The store opened with featured partners that included Khan Academy (Code Tutor), Canva, AllTrails, Consensus, Books, and the CK-12 Flexi AI tutor.
At launch the store was restricted to paying customers on the ChatGPT Plus, Team, or Enterprise plans. The new ChatGPT Team plan was announced the same day.
On May 13, 2024, OpenAI introduced GPT-4o and made several previously paid features available to free users, including the ability to browse and use GPTs from the store. Free users can run Custom GPTs but still cannot build their own; building remains a paid feature. Free users also hit a daily message cap, after which the conversation falls back to a smaller model.
Late January 2024 brought "GPT mentions," a feature that lets a Plus or Enterprise user invoke a Custom GPT inside a normal chat by typing @ followed by the GPT's name. The selected GPT then answers from inside the same conversation, which makes it easier to chain different specialists. OpenAI has also added a builder profile page, version history, and the ability to require sign-in for private GPTs.
The GPT Builder is the editor that runs inside ChatGPT at chatgpt.com/gpts/editor. It has two tabs:
Most serious builders end up using Configure for the system prompt because the conversational drafts tend to be generic.
A builder can attach up to 20 files per GPT, each up to 512 MB. Supported formats include PDF, DOCX, TXT, CSV, code files, and common image types. When a user asks a question, ChatGPT searches the attached files using retrieval and feeds relevant chunks back into the prompt. Retrieval is not perfect: for very large documents, asking about a specific page or table can still produce wrong or hallucinated answers, which is why many builders chunk their content manually before uploading.
Knowledge files are not part of the public profile. A user can read excerpts that the GPT cites, but the raw files are intended to be private to the builder. In practice, security researchers have shown that a determined user can sometimes coax a GPT into leaking the contents of its instructions or knowledge files through prompt injection, which is one reason builders are warned not to upload truly sensitive material.
The builder can toggle three built-in capabilities:
These are the same tools available in regular ChatGPT, but a Custom GPT can be locked to just the ones the builder picks. A writing GPT, for example, can disable code execution to keep the conversation focused.
Actions are how a Custom GPT talks to the rest of the internet. The builder pastes in an OpenAPI 3 schema describing a third-party API, plus an authentication setup (API key, OAuth, or none). At runtime ChatGPT uses function calling to decide when an API call is needed, fills in the JSON, sends the request, and reads the response back into the conversation.
Actions replaced the earlier ChatGPT "plugins" system, which OpenAI sunset in April 2024. Anything plugins could do, Actions can do, with the difference that Actions live inside a specific GPT rather than a global marketplace. Common Actions hook up Zapier, Notion, calendars, e-commerce backends, internal company APIs, and search providers.
A crucial privacy detail: when a user's message triggers an Action, the relevant parts of that message and the response are sent to the third party. OpenAI does not audit or control how those third parties handle the data, and the user sees a permission prompt the first time a new Action runs.
Conversation starters are up to four short example prompts that appear as buttons when a user opens the GPT. They are not part of the instructions; they are user-facing suggestions that help people figure out what the GPT is for. A good starter is the difference between a GPT that feels intuitive and one that feels like a blank box.
Every GPT has a public-looking profile with a name, description, builder name, conversation starters, and a generated icon. Builders choose visibility:
Verified builder names appear on the profile. Builders can use their real name or a verified domain (for example, a company can show khanacademy.org next to its GPTs).
The GPT Store is the public directory for Custom GPTs. It opened on January 10, 2024 inside ChatGPT and is reachable from the sidebar. The home page features a curated "GPTs of the week," weekly editorial picks, and a community leaderboard sorted by usage in each category.
A published GPT must comply with OpenAI's usage policies, must declare its builder, and cannot impersonate a real person or organization without consent. OpenAI runs automated and human moderation against the store, although enforcement has been uneven (see the criticism section).
The GPT Store organizes Custom GPTs under a fixed list of top-level categories. The aiwiki collection mirrors that structure:
See also: DALL-E Custom GPTs
See also: Writing Custom GPTs
See also: Productivity Custom GPTs
| Custom GPT | Image | Description | Knowledge | Actions | Link |
|---|---|---|---|---|---|
| 22.500 Best Custom GPTs | * | Search all public GPTs in one place. Find the best Custom ChatGPTs tailored to your needs. Every day, hundreds of new popular GPTs join our ranks. Discover GPT Store's best! | Yes | Https://chat.openai.com//g/g-RuhDS8mbd-22-500-best-custom-gpts | |
| GPT Public Directory |  | A directory assistant for finding and registering GPTs. With 11,000+ GPTs Available! | Https://chat.openai.com//g/g-tQBmTaWqj-gpt-public-directory |
See also: Research & Analysis Custom GPTs
See also: Programming Custom GPTs
| Custom GPT | Image | Description | Knowledge | Actions | Link |
|---|---|---|---|---|---|
| WebGPT | * | ChatGPT with unbiased access to the Web, can build products using No-Code playgrounds, and use API's. Powered by Web Requests. | Yes | Https://chat.openai.com//g/g-9MFRcOPwQ-webgpt | |
| Code Copilot (GPT) |  | Code Smarter, Build Faster, With the Expertise of a 10x Programmer by Your Side. | Yes | Yes | Https://chat.openai.com//g/g-2DQzU5UZl-code-copilot |
See also: Education Custom GPTs
See also: Lifestyle Custom GPTs
See also: Custom GPTs By ChatGPT
The GPTs OpenAI featured at the store launch on January 10, 2024 included:
| GPT | Builder | Purpose |
|---|---|---|
| AllTrails | AllTrails | Personalized hiking trail recommendations |
| Khan Academy Code Tutor | Khan Academy | Beginner coding tutor |
| Canva | Canva | Generates presentation and social media designs |
| Consensus | Consensus | Searches and summarizes 200 million academic papers |
| Books | OpenAI | Reading recommendations and book discussions |
| CK-12 Flexi | CK-12 Foundation | AI tutor for math and science |
Beyond the curated set, the most-used public categories tend to be image generation (Image, DALL-E variants), writing assistants, coding helpers, and research tools. Internally, the pharmaceutical company Moderna has said it built around 750 GPTs for staff use on top of ChatGPT Enterprise.
At launch OpenAI said it would pay builders based on user engagement, with details to follow. In March 2024 it opened a pilot "GPT builder revenue program" to a small group of US-based builders, paying out based on how much paying ChatGPT users engaged with a builder's GPTs each week. Reporting from independent builders suggests the formula has a minimum engagement threshold (around 25 unique paying-user conversations per week is commonly cited), under which a GPT earns nothing.
The program is US-only at the time of this writing. Builders outside the United States can still publish GPTs and rank in the store, but they cannot collect payouts. OpenAI has said it intends to expand the program internationally but has not committed to a public schedule. Top earners have reported payouts ranging from a few hundred to a few thousand dollars a month; most listed GPTs earn nothing.
GPT conversations on consumer plans are subject to the standard ChatGPT data policy: by default, conversations may be used to improve OpenAI's models unless the user opts out in settings. Enterprise and Team plans are not used for training and offer admin controls.
Four specific privacy points are worth knowing:
In 2024, Palo Alto Networks and several academic groups published reports detailing instruction-leak and file-extraction attacks against widely shared GPTs, recommending that organizations treat any content uploaded to a public Custom GPT as eventually disclosable.
The GPT Store grew faster than OpenAI's moderation, and the gap shows. A March 2024 TechCrunch investigation found the store full of GPTs that:
A Gizmodo follow-up in September 2024 documented more than 100 active GPTs that violated content policy, ranging from sexual content generators to unlicensed medical and legal advisors. OpenAI has tightened policies and removed offending GPTs in batches, but enforcement remains reactive rather than proactive.
Teachers and authors have also flagged GPTs that appear to reproduce copyrighted textbook material on demand, sometimes by uploading scanned chapters as knowledge files. OpenAI takes these down on report, although by then a user has typically already used the resource.
A separate complaint from builders concerns the store's discovery model. Rankings are opaque, the leaderboard heavily favors GPTs that arrived early, and clones of popular GPTs (sometimes with near-identical names) regularly appear and confuse users.
Custom GPTs are the most visible example of a broader category of "configurable assistants," but they are not the only one. The clearest comparison points are inside the Anthropic and Google ecosystems.
| Feature | Custom GPTs (OpenAI) | Claude Projects (Anthropic) | Gemini Gems (Google) |
|---|---|---|---|
| Launched | November 2023 | June 2024 | August 2024 |
| Built on | GPT-4 / GPT-4o family | Claude 3 / 3.5 family | Gemini 1.5 / 2.0 |
| Custom instructions | Yes | Yes | Yes |
| Knowledge files | 20 files up to 512 MB each, retrieval-based | Up to 200K-token context loaded directly | Files plus retrieval |
| Web browsing | Yes | No (native), tool-based since 2025 | Yes |
| Code execution | Yes (Python sandbox) | Limited; Artifacts since 2024 | Yes |
| Image generation | Yes (DALL-E) | No native | Yes (Imagen) |
| Third-party API actions | Yes (OpenAPI) | Through MCP since late 2024 | Through extensions |
| Public marketplace | Yes (GPT Store) | No public store | No public store |
| Revenue sharing | US builders only, engagement-based | None | None |
| Required plan for building | Plus, Team, Enterprise, Edu | Pro, Team, Enterprise | Advanced, Workspace |
The headline trade-off is reach versus context. Claude Projects load all uploaded files directly into Claude's long context window, which is usually a more reliable way to answer questions about specific pages or passages. Custom GPTs use retrieval, which scales to bigger document collections but can miss the exact passage on any given query. On the other side, the GPT Store gives builders a real audience and the prospect of payment; Claude Projects and Gemini Gems are useful but stay private to the builder or the builder's team.
For regulated work (legal, medical, internal codebases), many teams pick Claude Projects because Anthropic does not train on user data by default and the privacy story is simpler. For consumer-facing tools or anything that needs DALL-E, web search, or a Python sandbox in one place, Custom GPTs remain the easiest path.
OpenAI has also been pulling related concepts into ChatGPT itself: custom instructions, memory, and Projects in ChatGPT (a folder-style workspace introduced in late 2024 that bundles files, instructions, and chats without publishing a public GPT). These features overlap with Custom GPTs and are starting to absorb the simpler use cases that would once have been solved by spinning up a private GPT.