Documentation 2026-05-07 21:57 UTC to 2026-05-12 18:57 UTC

28 - Click on "Create app" next to **Advanced settings** and create an app for your remote MCP server. It will appear in the composer's "Developer Mode" tool later during conversations. The "Create app" button will only show if you are in Developer mode.28 - Click on "Create app" next to **Advanced settings** and create an app for your remote MCP server. It will appear in the composer's "Developer Mode" tool later during conversations. The "Create app" button will only show if you are in Developer mode.

29 - Supported MCP protocols: SSE and streaming HTTP.29 - Supported MCP protocols: SSE and streaming HTTP.

30 - Authentication supported: OAuth, No Authentication, and Mixed Authentication30 - Authentication supported: OAuth, No Authentication, and Mixed Authentication

31 - For OAuth, if static credentials are provided, then they will be used. Otherwise, dynamic client registration will be used to create the credentials.31 - For OAuth, if static credentials are provided, then they will be used. Otherwise, ChatGPT can use Client ID Metadata Documents when the authorization server advertises support and the connector creator chooses CIMD. CIMD supports public-client token exchange (`none`) and signed client assertion token exchange (`private_key_jwt`). ChatGPT can also use DCR when configured.

32 - Mixed authentication is supporting Oauth and No Authentication. This means the initialize and list tools APIs are no auth, and tools will be Oauth or Noauth based on the security schemes set on their tool metadata.32 - Mixed authentication supports OAuth and No Authentication. This means the initialize and list tools APIs use no auth, and tools use OAuth or no auth based on the security schemes set on their tool metadata.

33 - Created apps will show under "Drafts" in the app settings.33 - Created apps will show under "Drafts" in the app settings.

34- **Manage tools:** In app settings there is a details page per app. Use that to toggle tools on or off and refresh apps to pull new tools and descriptions from the MCP server.34- **Manage tools:** In app settings there is a details page per app. Use that to toggle tools on or off and refresh apps to pull new tools and descriptions from the MCP server.

35- **Use apps in conversations:** Choose **Developer mode** from the Plus menu and select the apps for the conversation. You may need to explore different prompting techniques to call the correct tools. For example:35- **Use apps in conversations:** Choose **Developer mode** from the Plus menu and select the apps for the conversation. You may need to explore different prompting techniques to call the correct tools. For example:

2 2 

3If you're not getting strong results with a fine-tuned model, consider the following iterations on your process.3If you're not getting strong results with a fine-tuned model, consider the following iterations on your process.

4 4 

5OpenAI is winding down the fine-tuning platform. The platform is no longer

6 accessible to new users, but existing users of the fine-tuning platform will

7 be able to create training jobs for the coming months.

8 <br />

9 All fine-tuned models will remain available for inference until their base

10 models are [deprecated](https://developers.openai.com/api/docs/deprecations). The full timeline is

11 [here](https://developers.openai.com/api/docs/deprecations).

12 

5### Iterating on data quality13### Iterating on data quality

6 14 

7Below are a few ways to consider improving the quality of your training data set:15Below are a few ways to consider improving the quality of your training data set:

64 65 

65## Fine-tune a model66## Fine-tune a model

66 67 

68OpenAI is winding down the fine-tuning platform. The platform is no longer

69 accessible to new users, but existing users of the fine-tuning platform will

70 be able to create training jobs for the coming months.

71 <br />

72 All fine-tuned models will remain available for inference until their base

73 models are [deprecated](https://developers.openai.com/api/docs/deprecations). The full timeline is

74 [here](https://developers.openai.com/api/docs/deprecations).

75 

67OpenAI models are already pre-trained to perform across a broad range of subjects and tasks. Fine-tuning lets you take an OpenAI base model, provide the kinds of inputs and outputs you expect in your application, and get a model that excels in the tasks you'll use it for.76OpenAI models are already pre-trained to perform across a broad range of subjects and tasks. Fine-tuning lets you take an OpenAI base model, provide the kinds of inputs and outputs you expect in your application, and get a model that excels in the tasks you'll use it for.

68 77 

69Fine-tuning can be a time-consuming process, but it can also enable a model to consistently format responses in a certain way or handle novel inputs. You can use fine-tuning with [prompt engineering](https://developers.openai.com/api/docs/guides/text) to realize a few more benefits over prompting alone:78Fine-tuning can be a time-consuming process, but it can also enable a model to consistently format responses in a certain way or handle novel inputs. You can use fine-tuning with [prompt engineering](https://developers.openai.com/api/docs/guides/text) to realize a few more benefits over prompting alone:

120| `low` | You need responsiveness plus basic reasoning. | Customer support, order lookup, simple policy questions. |120| `low` | You need responsiveness plus basic reasoning. | Customer support, order lookup, simple policy questions. |

121| `medium` | The assistant must reason through multi-step tasks. | Technical support, diagnostics, complex routing. |121| `medium` | The assistant must reason through multi-step tasks. | Technical support, diagnostics, complex routing. |

122| `high` | Deeper reasoning materially improves success. | High-precision workflows, escalation decisions, tasks with constraints. |122| `high` | Deeper reasoning materially improves success. | High-precision workflows, escalation decisions, tasks with constraints. |

123| `xhigh` | Maximum reasoning is worth added latency and cost. | Complex planning, critical triage, high-stakes tool orchestration. |

123 124 

124Beyond the API setting, steer the model on when and how much to reason.125Beyond the API setting, steer the model on when and how much to reason.

125 126 

2 2 

3Reinforcement fine-tuning (RFT) adapts an OpenAI reasoning model with a feedback signal you define. Like [supervised fine-tuning](https://developers.openai.com/api/docs/guides/supervised-fine-tuning), it tailors the model to your task. The difference is that instead of training on fixed “correct” answers, it relies on a programmable grader that scores every candidate response. The training algorithm then shifts the model’s weights, so high-scoring outputs become more likely and low-scoring ones fade.3Reinforcement fine-tuning (RFT) adapts an OpenAI reasoning model with a feedback signal you define. Like [supervised fine-tuning](https://developers.openai.com/api/docs/guides/supervised-fine-tuning), it tailors the model to your task. The difference is that instead of training on fixed “correct” answers, it relies on a programmable grader that scores every candidate response. The training algorithm then shifts the model’s weights, so high-scoring outputs become more likely and low-scoring ones fade.

4 4 

5OpenAI is winding down the fine-tuning platform. The platform is no longer

6 accessible to new users, but existing users of the fine-tuning platform will

7 be able to create training jobs for the coming months.

8 <br />

9 All fine-tuned models will remain available for inference until their base

10 models are [deprecated](https://developers.openai.com/api/docs/deprecations). The full timeline is

11 [here](https://developers.openai.com/api/docs/deprecations).

12 

5<br />13<br />

6 14 

7<table>15<table>

2 2 

3[Reinforcement fine-tuning](https://developers.openai.com/api/docs/guides/reinforcement-fine-tuning) (RFT) provides a way to improve your model's performance at specific tasks. The task must be clear and have verifiable answers.3[Reinforcement fine-tuning](https://developers.openai.com/api/docs/guides/reinforcement-fine-tuning) (RFT) provides a way to improve your model's performance at specific tasks. The task must be clear and have verifiable answers.

4 4 

5OpenAI is winding down the fine-tuning platform. The platform is no longer

6 accessible to new users, but existing users of the fine-tuning platform will

7 be able to create training jobs for the coming months.

8 <br />

9 All fine-tuned models will remain available for inference until their base

10 models are [deprecated](https://developers.openai.com/api/docs/deprecations). The full timeline is

11 [here](https://developers.openai.com/api/docs/deprecations).

12 

5## When to use reinforcement fine-tuning13## When to use reinforcement fine-tuning

6 14 

7Agentic workflows are designed to make decisions that are both correct and verifiable. RFT can help by providing explicit rubrics and using code‑based or LLM‑based graders to measure functional success, factual accuracy, or policy compliance.15Agentic workflows are designed to make decisions that are both correct and verifiable. RFT can help by providing explicit rubrics and using code‑based or LLM‑based graders to measure functional success, factual accuracy, or policy compliance.

12 customUserLocationExampleCoarse,12 customUserLocationExampleCoarse,

13 customUserLocationExampleCoarseChat,13 customUserLocationExampleCoarseChat,

14 listSourcesExample,14 listSourcesExample,

15 returnTokenBudgetExample,

15 searchContextSize,16 searchContextSize,

16} from "./web-search-examples";17} from "./web-search-examples";

17 18 

28## Choose an integration29## Choose an integration

29 30 

30| Use case | Recommended path | Notes |31| Use case | Recommended path | Notes |

31| --------------------------------------------- | --------------------------------------------- | ------------------------------------------------------------------------------------- |32| --------------------------------------------- | --------------------------------------------- | ----------------------------------------------------------------------------------------------------------- |

32| New web search integration | Responses API with `web_search` and `gpt-5.5` | Supports hosted web search controls such as filters, sources, and live-access control |33| New web search integration | Responses API with `web_search` and `gpt-5.5` | Supports hosted web search controls such as filters, sources, live-access control, and longer research runs |

33| Existing Chat Completions search integration | Chat Completions with `gpt-5-search-api` | Use this only when you need to preserve a Chat Completions integration |34| Existing Chat Completions search integration | Chat Completions with `gpt-5-search-api` | Use this only when you need to preserve a Chat Completions integration |

34| Multi-step research or long-running reporting | `gpt-5.5` with `high` or `xhigh` reasoning | Use background mode for reports that can take several minutes |35| Multi-step research or long-running reporting | `gpt-5.5` with `high` or `xhigh` reasoning | Use background mode for reports that can take several minutes |

35 36 

36Using the [Responses API](https://developers.openai.com/api/docs/api-reference/responses), you can enable web search by configuring it in the `tools` array in an API request to generate content. Like any other tool, the model can choose to search the web or not based on the content of the input prompt.37Using the [Responses API](https://developers.openai.com/api/docs/api-reference/responses), you can enable web search by configuring it in the `tools` array in an API request to generate content. Like any other tool, the model can choose to search the web or not based on the content of the input prompt.

37 38 

38For new Responses API integrations, use `{ "type": "web_search" }`. The earlier `web_search_preview` tool remains available for legacy integrations, but it does not support newer controls such as `filters` and `external_web_access`.39For new Responses API integrations, use `{ "type": "web_search" }`. The earlier `web_search_preview` tool remains available for legacy integrations, but it does not support newer controls such as `filters`, `external_web_access`, and `return_token_budget`.

39 40 

40## Output and citations41## Output and citations

41 42 

97## Migrating from legacy web search98## Migrating from legacy web search

98 99 

99| If you use | Recommended path | Notes |100| If you use | Recommended path | Notes |

100| ------------------------------------------------------- | ------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------ |101| ------------------------------------------------------- | ------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------- |

101| `web_search_preview` in Responses | Migrate to `web_search` | `web_search` supports newer controls such as `filters` and `external_web_access` |102| `web_search_preview` in Responses | Migrate to `web_search` | `web_search` supports newer controls such as `filters`, `external_web_access`, and `return_token_budget` |

102| `gpt-4o-search-preview` or `gpt-4o-mini-search-preview` | Migrate to Responses `web_search`, or use `gpt-5-search-api` if you must stay on Chat Completions | The preview search models are deprecated and shut down on 2026-07-23 |103| `gpt-4o-search-preview` or `gpt-4o-mini-search-preview` | Migrate to Responses `web_search`, or use `gpt-5-search-api` if you must stay on Chat Completions | The preview search models are deprecated and shut down on 2026-07-23 |

103| Chat Completions search integrations | Use `gpt-5-search-api`, or migrate to Responses `web_search` for more tool controls and optional search | Chat Completions search models always search before responding; Responses search is a tool |104| Chat Completions search integrations | Use `gpt-5-search-api`, or migrate to Responses `web_search` for more tool controls and optional search | Chat Completions search models always search before responding; Responses search is a tool |

104 105 

108 109 

109 110 

110 111 

112## Run longer web research

113 

114`return_token_budget` controls how much web search result content the tool can return during a Responses API search run with GPT-5+ reasoning models. Keep the default for most requests. Set it to `unlimited` only for high-effort research or evaluation runs that need to inspect many pages and might otherwise stop at the standard returned-token cap.

115 

116Use `unlimited` selectively because it can increase latency and cost. For long-running multi-search tasks, use background mode (`background: true`) so the request can keep running asynchronously and you can retrieve the final response later.

117 

118| Value | Behavior |

119| ----------- | ---------------------------------------------------------------------------------------------------------------------------- |

120| `default` | Uses the standard returned-token budget for web search results. This is the same behavior as omitting `return_token_budget`. |

121| `unlimited` | Removes the default returned-token budget for the web search run. |

122 

123This parameter applies only to the hosted Responses API `web_search` tool with GPT-5+ reasoning web search. It does not change the search context window, and it does not apply to non-reasoning web search, legacy Search API paths, container web search, Chat Completions search models, or `web_search_preview`. Only `default` and `unlimited` are supported values; `null`, numbers, and other strings are rejected.

124 

125 

126 

111## Domain filtering127## Domain filtering

112 128 

113Domain filtering in web search lets you limit results to a specific set of domains. With the `filters` parameter you can configure up to 100 `allowed_domains` or up to 100 `blocked_domains`. When formatting domains, omit the HTTP or HTTPS prefix. For example, use `openai.com` instead of `https://openai.com/`. This approach also includes subdomains in the search. Note that domain filtering is only available in the Responses API with the `web_search` tool.129Domain filtering in web search lets you limit results to a specific set of domains. With the `filters` parameter you can configure up to 100 `allowed_domains` or up to 100 `blocked_domains`. When formatting domains, omit the HTTP or HTTPS prefix. For example, use `openai.com` instead of `https://openai.com/`. This approach also includes subdomains in the search. Note that domain filtering is only available in the Responses API with the `web_search` tool.

146 162 

147#### Chat Completions API163#### Chat Completions API

148 164 

149The Chat Completions API supports only specialized search models for web search. These models do not support Responses API `web_search` features such as domain filters, complete source lists, live-access control.165The Chat Completions API supports only specialized search models for web search. These models do not support Responses API `web_search` features such as domain filters, complete source lists, live-access control, and returned-token budget control.

150 166 

151| Model | Context window | Limitation |167| Model | Context window | Limitation |

152| ---------------------------- | -------------: | -------------------------------------------------------------------------------------------------------------------------------------------- |168| ---------------------------- | -------------: | -------------------------------------------------------------------------------------------------------------------------------------------- |

171- Web search does not support [`gpt-5`](https://developers.openai.com/api/docs/models/gpt-5) with `minimal` reasoning.187- Web search does not support [`gpt-5`](https://developers.openai.com/api/docs/models/gpt-5) with `minimal` reasoning.

172- [`gpt-5.4`](https://developers.openai.com/api/docs/models/gpt-5.4) with reasoning effort set to `none` may produce lower-quality results.188- [`gpt-5.4`](https://developers.openai.com/api/docs/models/gpt-5.4) with reasoning effort set to `none` may produce lower-quality results.

173- Responses API web search uses the underlying model's tiered rate limits.189- Responses API web search uses the underlying model's tiered rate limits.

174- `web_search_preview` does not support `filters` and ignores `external_web_access`.190- `web_search_preview` does not support `filters` or `return_token_budget`, and ignores `external_web_access`.

175- With `tool_choice: "auto"`, search is optional. Use `tool_choice: "required"` or a specific web search tool choice when search must run.191- With `tool_choice: "auto"`, search is optional. Use `tool_choice: "required"` or a specific web search tool choice when search must run.

176 192 

177## Usage notes193## Usage notes

2 2 

3Vision fine-tuning uses image inputs for [supervised fine-tuning](https://developers.openai.com/api/docs/guides/supervised-fine-tuning) to improve the model's understanding of image inputs. This guide will take you through this subset of SFT, and outline some of the important considerations for fine-tuning with image inputs.3Vision fine-tuning uses image inputs for [supervised fine-tuning](https://developers.openai.com/api/docs/guides/supervised-fine-tuning) to improve the model's understanding of image inputs. This guide will take you through this subset of SFT, and outline some of the important considerations for fine-tuning with image inputs.

4 4 

5OpenAI is winding down the fine-tuning platform. The platform is no longer

6 accessible to new users, but existing users of the fine-tuning platform will

7 be able to create training jobs for the coming months.

8 <br />

9 All fine-tuned models will remain available for inference until their base

10 models are [deprecated](https://developers.openai.com/api/docs/deprecations). The full timeline is

11 [here](https://developers.openai.com/api/docs/deprecations).

12 

5<br />13<br />

6 14 

7<table>15<table>

416 416 

417### Handle authentication417### Handle authentication

418 418 

419As someone building a custom remote MCP server, authorization and authentication help you protect your data. We recommend using OAuth and [dynamic client registration](https://modelcontextprotocol.io/specification/2025-03-26/basic/authorization#2-4-dynamic-client-registration). For ChatGPT app auth requirements, see [Authentication](https://developers.openai.com/apps-sdk/build/auth). For protocol details, read the [MCP user guide](https://modelcontextprotocol.io/docs/concepts/transports#authentication-and-authorization) or the [authorization specification](https://modelcontextprotocol.io/specification/2025-03-26/basic/authorization).419As someone building a custom remote MCP server, authorization and authentication help you protect your data. We recommend using OAuth with [Client ID Metadata Documents](https://modelcontextprotocol.io/specification/2025-11-25/basic/authorization#client-id-metadata-documents) for client registration when your authorization server supports CIMD and the connector creator chooses it. ChatGPT supports CIMD with public-client token exchange (`none`) or signed client assertion token exchange (`private_key_jwt`). Dynamic client registration remains supported when configured. For ChatGPT app auth requirements, see [Authentication](https://developers.openai.com/apps-sdk/build/auth). For protocol details, read the [MCP user guide](https://modelcontextprotocol.io/docs/concepts/transports#authentication-and-authorization) or the [authorization specification](https://modelcontextprotocol.io/specification/2025-11-25/basic/authorization).

420 420 

421If you connect your custom remote MCP server in ChatGPT as an app, users in your workspace will get an OAuth flow to your application.421If you connect your custom remote MCP server in ChatGPT as an app, users in your workspace will get an OAuth flow to your application.

422 422