OpenAI Multimodal Responses API

Authentication header. Uses a Bearer token, e.g. Bearer YOUR_API_KEY.

Request content type, must be application/json.

Model name, e.g. gpt-4o or a reasoning model.

Input to the model. Can be a plain text string (equivalent to a text input with the user role), or an array of one or more input items containing different content types such as { "type": "input_text", ... } and { "type": "input_image", ... }.

A system (or developer) message inserted into the model's context. When used together with previous_response_id, the instructions from a previous response are not carried over to the next response.

Reasoning configuration, only applicable to reasoning models (o-series and gpt-5 series).

Text output configuration.

An array of tools the model may call while generating a response. You can specify which tool to use via tool_choice. Three categories are supported: built-in tools (web_search, file_search, computer_use / computer_use_preview, code_interpreter, image_generation, etc.), MCP tools ({ "type": "mcp", "server_label": ..., ... }), and function tools ({ "type": "function", "name": ..., "description": ..., "parameters": ...(JSON Schema), "strict": ... }). Other upstream-compatible tools can also be forwarded.

How the model should select which tool(s) to use. String values: none (never call a tool), auto (the model decides, default), required (must call one or more tools). Or pass an object to force a specific tool, e.g. { "type": "function", "name": "my_function" }, { "type": "mcp", "server_label": "..." }, { "type": "allowed_tools", "mode": "auto" | "required", "tools": [...] }, or a hosted tool type such as { "type": "file_search" }, { "type": "web_search_preview" }, { "type": "computer_use_preview" }, { "type": "image_generation" }, { "type": "code_interpreter" }.

Whether to allow the model to run tool calls in parallel. Defaults to true.

An upper bound for the number of tokens that can be generated for a response, including visible output tokens and reasoning tokens. Minimum value is 16. Explicitly passing 0 will be preserved and forwarded to supported upstreams.

The maximum number of total calls to built-in tools that can be processed in a response, counted across all built-in tool calls. Further tool call attempts by the model beyond this limit will be ignored.

Sampling temperature, between 0 and 2. Defaults to 1. Higher values make the output more random. It is generally recommended to alter this or top_p, but not both.

Nucleus sampling parameter, between 0 and 1. Defaults to 1. For example, 0.1 means only the tokens comprising the top 10% probability mass are considered. It is generally recommended to alter this or temperature, but not both.

An integer between 0 and 20 specifying the maximum number of most likely tokens to return at each token position, each with an associated log probability.

Whether to stream the response with server-sent events (SSE). Defaults to false.

Options for streaming responses. Only set this when stream: true.

Whether to run the model response in the background. Defaults to false. Background responses can be retrieved later.

Whether the upstream stores the generated response for later retrieval via API. Defaults to true. This field is allowed to be forwarded by default and can be disabled by channel settings.

Key-value pairs attached to the response object. Up to 16 pairs; keys up to 64 characters, values up to 512 characters.

Additional output data to include in the model response. Supported values: web_search_call.results, web_search_call.action.sources, code_interpreter_call.outputs, computer_call_output.output.image_url, file_search_call.results, message.input_image.image_url, message.output_text.logprobs, reasoning.encrypted_content (carries encrypted reasoning content for multi-turn conversations in stateless scenarios, e.g. store: false or zero data retention organizations).

(Deprecated) The truncation strategy: auto (if the input exceeds the context window, the model truncates by dropping items from the beginning of the conversation) or disabled (fail with a 400 error, default).

The unique ID of the previous response to the model, used to create multi-turn conversations. Cannot be used in conjunction with conversation. Can be used for context continuation when supported by the upstream.

The conversation this response belongs to. Can be a conversation ID string or a { "id": "conv_..." } object. Items from this conversation are prepended to the input items for this request, and the input/output items of this response are automatically added to the conversation afterwards. Cannot be used in conjunction with previous_response_id.

Reference to a prompt template and its variables.

Used to cache responses for similar requests and optimize cache hit rates. Replaces the deprecated user field.

Options for prompt caching (supported for gpt-5.6 and later models). Contains ttl (minimum lifetime of cache entries, defaults to 30m) and mode (implicit, the default, automatically creates an implicit cache breakpoint; explicit uses only explicit breakpoints).

A stable identifier used to help detect users of your application that may be violating usage policies. Maximum 64 characters. It is recommended to hash the username or email address to avoid sending identifying information.

The processing tier used for serving the request: auto (use the Project settings, default), default (standard pricing and performance), flex (Flex Processing), scale, priority / fast (priority processing).

(Deprecated) A stable identifier for your end-users. Use safety_identifier and prompt_cache_key instead.

Configuration for running moderation on the input and output of this response. Contains model (the moderation model, e.g. omni-moderation-latest, required) and policy (moderation policy for input/output, score or block).

Context management configuration for this request, controlling how oversized contexts are handled. The exact structure depends on the upstream implementation.