Custom GPT

A custom GPT, branded by OpenAI as a "GPT", is a version of ChatGPT configured for a particular purpose. Its builder can supply instructions, conversation starters, reference files, selected built-in capabilities, and either connected apps or custom API actions. The result is used inside ChatGPT; it is not a standalone model or an application embedded through the OpenAI API.^[1][3]

OpenAI announced GPTs on November 6, 2023. The company presented them as a no-code way to tailor ChatGPT for personal, organizational, or public use. It opened the GPT Store on January 10, 2024 and said that users had created more than 3 million GPTs during the first two months.^[1][2]

Any signed-in ChatGPT user can use a GPT that they are allowed to access. Creating or editing one requires a paid plan, and managed workspaces can impose additional role and sharing restrictions.^[3] This article covers the configured assistant itself. The marketplace is covered separately at GPT Store, and the earlier extension system is covered at ChatGPT plugins.

Definition and scope

A custom GPT is a saved ChatGPT configuration. OpenAI's documentation lists six main elements: instructions, conversation starters, uploaded knowledge, built-in capabilities, apps, and actions.^[3] Instructions apply to every conversation with that GPT. Knowledge files provide reference material. Capabilities add OpenAI-provided tools, while apps and actions connect the GPT to outside services.^[4][5]

This configuration is distinct from the underlying large language model. A builder may recommend a model, but users can switch models when alternatives are available. If the recommended model is unavailable, ChatGPT may select a similar model automatically.^[4] OpenAI also states that GPTs do not use a user's saved memory, custom instructions, or previous conversations, so each GPT conversation starts without those forms of personal context.^[3]

GPTs are designed to run within ChatGPT. OpenAI explicitly distinguishes them from developer-built assistants that use its API inside another website or service. A GPT cannot itself be used as an embeddable ChatGPT component.^[3]

History

Launch in November 2023

OpenAI introduced GPTs at its first DevDay conference on November 6, 2023. The launch examples included assistants for learning board games, teaching mathematics, and designing stickers. OpenAI also made examples from Canva and Zapier available to ChatGPT Plus and Enterprise users.^[1]

The same event introduced the Assistants API. OpenAI said that the API was built on capabilities also used by GPTs, including custom instructions, Code Interpreter, retrieval, and function calling. The products nevertheless served different interfaces: GPTs provided a no-code experience inside ChatGPT, while the API supported integrations built by developers.^[7]

GPT Store launch in January 2024

OpenAI initially expected to open the GPT Store in November 2023. A December memo reported by Axios moved the launch to early 2024 and referred only to unspecified "unexpected things" that had occupied the company. The available record does not establish a more specific cause.^[8] The Store opened on January 10, 2024 to ChatGPT Plus, Team, and Enterprise users.^[2]

At launch, the Store organized GPTs into categories including writing, research, programming, education, lifestyle, and image generation. OpenAI also introduced featured listings and usage-based rankings.^[2]

Free access and model changes

OpenAI began extending access to GPTs created by other users to the Free tier alongside the May 2024 release of GPT-4o.^[9] By July 2026, OpenAI's documentation stated that all signed-in ChatGPT users could use GPTs, while creation and editing remained paid features.^[3]

Free-tier access is governed by usage limits. When a Free user reaches the model limit that applies to GPT use, access to GPTs pauses until that limit resets. This differs from an ordinary chat switching to a smaller model after a limit is reached.^[10]

Custom GPTs also follow changes to ChatGPT's model lineup. OpenAI retired GPT-4o, GPT-4.1, GPT-4.1 mini, o4-mini, and the Instant and Thinking versions of GPT-5 from ChatGPT on February 13, 2026. It retired GPT-5.1 Instant, Thinking, and Pro on March 11. Business, Enterprise, and Edu customers retained GPT-4o in custom GPTs until April 3, after which OpenAI described it as fully retired across plans. GPTs configured with retired models were moved to the closest available GPT-5.3 Instant or GPT-5.4 Thinking or Pro equivalent.^[11]

Building and configuration

Building and editing are available through the web version of ChatGPT. Mobile apps can use GPTs but cannot build them. The editor supports a conversational builder, which drafts a configuration from a description, and a direct configuration view for editing individual fields.^[4]

Scroll sideways for more →

The editor supports common document, spreadsheet, image, text, and code formats. Accepted types may vary by model, and some formats require Code Interpreter and Data Analysis. OpenAI recommends text-forward files because complex layouts can make uploaded content harder for a GPT to use.^[4]

The distinction between instructions and knowledge is functional. Instructions describe how the GPT should behave. Knowledge files provide material that it can consult. OpenAI recommends putting rules and workflows in instructions rather than treating uploaded documents as behavior rules.^[4]

The editor saves drafts and maintains a version history. Builders can review and restore earlier versions, although restoring a version that used actions may require its authentication to be configured again.^[4]

Actions and apps

An action connects a GPT to an external API described by an OpenAPI document in JSON or YAML. The schema specifies servers, endpoints, parameters, and operation identifiers. The editor can accept a pasted schema, import one from a URL, or start from built-in examples.^[5]

Actions support three authentication modes:^[5]

Scroll sideways for more →

OAuth configuration requires a client identifier and secret, authorization and token URLs, scopes, and a token exchange method. The editor supplies a callback URL for completing the authorization flow.^[5]

A GPT can use apps or actions, but not both at the same time. OpenAI's July 2026 documentation also says that actions are unavailable in Pro mode, referring to a model mode rather than the ChatGPT Pro subscription. When a GPT has actions, the editor shows non-Pro models that support them.^[5]

Publicly shared GPTs with actions must provide a valid privacy-policy URL for each public action. Users may be asked to approve an action before it runs.^[5][6]

Sharing and permissions

The available sharing controls depend on the account and, in managed workspaces, administrator settings. OpenAI documented the following levels by June 2026:^[6]

Scroll sideways for more →

Personal accounts can share by public link or publish to the GPT Store. Direct sharing with named users or groups is limited to managed workspaces, where a GPT can be shared directly with up to 100 recipients. A group counts as one recipient.^[6]

Managed-workspace sharing can also grant different permissions. Can chat allows use of the GPT. Can view settings also allows duplication and inspection of the configuration. Can edit permits direct changes. Public-link and GPT Store sharing support only the chat permission, not permission to view settings.^[6]

Publishing to the GPT Store may require a Builder Profile with a name, a verified domain, or a supported verified social profile. A GPT may be ineligible if public publishing is disabled, an app connection is unsupported for the selected sharing method, an action lacks a privacy-policy URL, or a policy check blocks publication.^[6]

Access, privacy, and data handling

GPT builders cannot view individual conversations that users have with their GPTs. Whether OpenAI may use a conversation to improve its models depends on the user's plan and data controls. Business, Enterprise, and Edu data is not used for training by default. On consumer plans, including Free, Plus, Go, and Pro, use for training depends on whether the user has opted out.^[3]

Third-party integrations create a separate data path. When a GPT uses an app or external API, relevant portions of the user's input may be sent to that service. OpenAI says it does not audit or control how the third party stores or uses that data and advises users to use only GPTs whose integrations they trust.^[3]

If a builder cancels or downgrades the plan that enabled creation, the builder can continue using GPTs already created but cannot create new ones or edit the existing ones.^[3]

On ChatGPT web, a user can type @ in an ordinary conversation to select an accessible GPT. The next message goes to that GPT while the conversation retains its existing context. OpenAI's July 2026 documentation said this flow was unavailable in the iOS and Android apps.^[3]

GPT Store and monetization

The GPT Store is the public directory for eligible GPTs. At its January 2024 launch, OpenAI said that listings would be searchable and ranked, and that selected GPTs would be featured. A builder publishing publicly had to verify a Builder Profile and satisfy product and policy requirements.^[2]

OpenAI said at launch that it planned a builder revenue program in the first quarter of 2024, initially paying United States builders according to engagement. It did not publish a complete payment formula in that announcement.^[2] In October 2024, WIRED reported that revenue sharing had become a small, invitation-only pilot, that builders outside the United States were ineligible, and that several prominent builders interviewed for the article had received no direct Store payments. Those findings describe the program at that time and do not establish its later status.^[12]

Public review procedures also changed in wording over time. The January 2024 launch post described both human and automated review. OpenAI's June 2026 sharing documentation said a GPT may be checked automatically before broader sharing or Store publication, and provided an appeal path when publishing was blocked.^[2][6]

Organizational use and research

OpenAI promoted private organizational GPTs from the start, citing early work by Amgen, Bain, and Square for uses such as marketing material, customer support, and engineer onboarding.^[1] In an April 2024 case study, OpenAI reported that Moderna had created 750 GPTs within two months of adopting ChatGPT Enterprise and that 40% of weekly active users had created a GPT. These are company-reported figures from a promotional case study, not an independent adoption survey.^[13]

Independent studies found that public GPT use was uneven. A 2024 study by Benjamin Zhao and colleagues analyzed metadata for 334,348 GPTs collected through the Beetrove dataset. In that early-2024 snapshot, 28.5% had more than ten recorded conversations and 90.78% had no rating.^[16] A separate ten-month study by Zejun Zhang and colleagues reported that the top 1% of measured GPT apps accounted for 90% of daily interactions in its dataset. Both results depend on the stores, public metadata, and measurement periods studied, so they should not be read as current platform-wide usage statistics.^[15]

Security and limitations

Instructions and knowledge files should not be treated as secret merely because the editor does not expose them to ordinary users. In a study of 216 public GPTs, Jiahao Yu and colleagues reported extracting configured instructions from 210. Among the 24 tested GPTs that had uploaded files, the researchers reported retrieving all 24 files. The experiments were performed on the 2023-2024 product and therefore measure those versions of the service.^[14]

Zhang and colleagues later targeted the 10,000 most popular GPT apps in their dataset; 7,706 were publicly accessible when retrieved. They reported extracting system prompts from nearly 90% of those accessible apps. Among 2,638 apps with knowledge files, they obtained file names and download links from 88% and actual file contents from 12.7%. The researchers recommended against putting proprietary information in prompts and suggested external APIs as one way to keep sensitive data outside the GPT configuration.^[15]

Knowledge also does not guarantee a correct answer. OpenAI advises builders to test uploaded material in Preview and notes that text-forward files work better than complex layouts.^[4] A GPT remains dependent on the behavior of its underlying model, its instructions, the quality of its knowledge files, and any outside services it calls.

Moderation

OpenAI's launch materials said that Store listings were subject to policy and brand requirements and to automated and human review.^[2] A March 2024 TechCrunch investigation nevertheless found publicly listed GPTs that appeared to promote academic dishonesty, imitate public figures, or use copyrighted characters. In response, an OpenAI spokesperson said the company used automated systems, human review, and user reports, with possible consequences including warnings, sharing restrictions, and exclusion from the Store or monetization.^[17]

The investigation documents examples visible at that time rather than the prevalence of violations across the Store. Later sharing documentation described automatic policy checks, publishing restrictions, and appeals, but it did not promise that every public GPT would receive human review before publication.^[6]

See also

• ChatGPT
• GPT Store
• ChatGPT plugins
• OpenAI
• GPT-4o
• OpenAI Assistants API
• Function calling
• Large language model
• Prompt injection

References

1. ^OpenAI, "Introducing GPTs", 2023-11-06. openai.com/...introducing-gpts.
2. ^OpenAI, "Introducing the GPT Store", 2024-01-10. openai.com/...introducing-the-gpt-store.
3. ^OpenAI, "GPTs in ChatGPT", OpenAI Help Center, archived 2026-07-25. web.archive.org/...8554407-gpts-in-chatgpt.
4. ^OpenAI, "Creating and editing GPTs", OpenAI Help Center, archived 2026-07-09. web.archive.org/...8554397-creating-and-editing-gpts.
5. ^OpenAI, "Configuring actions in GPTs", OpenAI Help Center, archived 2026-07-25. web.archive.org/...513-configuring-actions-in-gpts.
6. ^OpenAI, "Sharing and publishing GPTs", OpenAI Help Center, archived 2026-06-23. web.archive.org/...878-sharing-and-publishing-gpts.
7. ^OpenAI, "New models and developer products announced at DevDay", 2023-11-06. openai.com/...veloper-products-announced-at-devday.
8. ^Ina Fried, "Scoop: OpenAI delays launch of custom GPT store until early 2024", Axios, 2023-12-01. axios.com/...i-delays-launch-custom-gpt-store-2024.
9. ^OpenAI, "Introducing GPT-4o and more tools to ChatGPT free users", 2024-05-13. openai.com/...gpt-4o-and-more-tools-to-chatgpt-free.
10. ^OpenAI, "ChatGPT Free Tier FAQ", OpenAI Help Center, archived 2026-06-22. web.archive.org/...9275245-chatgpt-free-tier-faq.
11. ^OpenAI, "Retiring GPT-4o and other ChatGPT models", OpenAI Help Center, archived 2026-07-25. web.archive.org/...gpt-4o-and-other-chatgpt-models.
12. ^Lauren Goode, "OpenAI's GPT Store Has Left Some Developers in the Lurch", WIRED, 2024-10-11. wired.com/...openai-gpt-store.
13. ^OpenAI, "Moderna", 2024-04-25. openai.com/...moderna.
14. ^Jiahao Yu, Yuhang Wu, Dong Shu, Mingyu Jin, Sabrina Yang, and Xingyu Xing, "Assessing Prompt Injection Risks in 200+ Custom GPTs", ICLR 2024 Workshop on Secure and Trustworthy Large Language Models, arXiv:2311.11538. arxiv.org/...2311.11538.
15. ^Zejun Zhang, Li Zhang, Xin Yuan, Anlan Zhang, Mengwei Xu, and Feng Qian, "A First Look at GPT Apps: Landscape and Vulnerability", arXiv:2402.15105, version 3, 2024-11-27. arxiv.org/...2402.15105.
16. ^Benjamin Zi Hao Zhao, Muhammad Ikram, and Mohamed Ali Kaafar, "GPTs Window Shopping: An Analysis of the Landscape of Custom ChatGPT Models", arXiv:2405.10547, 2024-05-17. arxiv.org/...2405.10547.
17. ^Kyle Wiggers, "OpenAI's chatbot store is filling up with spam", TechCrunch, 2024-03-20. techcrunch.com/...ot-store-is-filling-up-with-spam.