GPT All-Model API - Responses Reference

Model to call:

Model input: a plain string, or an array of input items.

The content of an input item supports two block types: input_text (text) and input_image (image):

"input": [
  {
    "role": "user",
    "content": [
      { "type": "input_text", "text": "What is in this image?" },
      {
        "type": "input_image",
        "image_url": "https://example.com/photo.png",
        "detail": "auto"
      }
    ]
  }
]

Image

• Pass the public URL of the image in image_url
• image_url must be a string; writing it as { "url": "..." } returns 400
• detail is a sibling of image_url (not nested inside it): auto (default) / low / high / original
• The image must be downloadable, otherwise 400 is returned

Tool results

• The array can also carry back tool result items from the previous turn, such as function_call_output

Note The block types on this API differ from those on the Chat Completions API (which uses text / image_url). They cannot be mixed; using the wrong ones returns 400.

System-level instructions, equivalent to inserting a system message at the very beginning of input. When continuing a conversation with previous_response_id, this parameter is not inherited from the previous turn and must be passed on every turn.

Whether to return a streaming response (SSE events, ending with response.completed). Default false.

Maximum number of tokens to generate (including reasoning tokens). When the limit is reached, status is incomplete.

GPT-6 Astra / Sol / Luna and GPT-6.1 Sol support up to 128,000 output tokens, including reasoning tokens.

Controls tool selection: "auto" (default) / "none" / "required", or an object pinning a specific tool, e.g. {"type": "web_search"}.

Maximum total number of tool calls allowed in this response (across all built-in tools).

Note When gpt-6-astra / gpt-6.1-sol / gpt-6-sol / gpt-6-luna uses image_generation and this parameter is omitted, a single request can generate up to 4 images; set it explicitly if you need more.

Whether the model may call multiple tools in parallel within one turn. Defaults to true.

GPT-6 Sol / Luna and GPT-6.1 Sol: support for this parameter has not been confirmed; omit it from basic requests.

Rules for existing models below exclude GPT-6 Sol / Luna and GPT-6.1 Sol:

Note gpt-6-astra, the gpt-5.6 family, and gpt-5.5 support setting this parameter to false; on gpt-5.4 / gpt-5.2 / gpt-5.1, it has no effect and always behaves as true.

The id of the previous response, used to chain multi-turn conversations without re-uploading the history.

Note Must be used together with store: true (the default). Responses have a retention period; once it expires the ID is no longer valid, and the request is handled as a new conversation without inheriting context. For scenarios with strict context-accuracy requirements, maintaining the full input history yourself is recommended.

Whether to store this response on the server; only stored responses can be referenced by previous_response_id. Defaults to true.

GPT-6 Sol / Luna and GPT-6.1 Sol: disabling retention with store: false has not been verified on the available channels. Acceptance of the field alone does not establish that the response was not retained.

Rules for existing models below exclude GPT-6 Sol / Luna and GPT-6.1 Sol:

Note gpt-6-astra, the gpt-5.6 family, and gpt-5.5 support setting this parameter to false; on gpt-5.4 / gpt-5.2 / gpt-5.1, it has no effect and always behaves as true. If you do not want responses stored, choose a model that supports turning it off.

Sampling temperature, ranging from 0 to 2. Lower values make the output more deterministic.

GPT-6: omit this parameter for gpt-6-astra and gpt-6.1-sol. For gpt-6-sol / gpt-6-luna, adjust it only with reasoning effort set to none; omit it at other effort levels. Omitting effort selects medium, not none.

Existing models: gpt-5.4 / gpt-5.2 / gpt-5.1 treat temperature: 0 as the default 1; use a positive value such as 0.01 for more deterministic output.

Nucleus sampling parameter, ranging from 0 to 1. Adjusting it together with temperature is not recommended.

GPT-6: omit this parameter for gpt-6-astra and gpt-6.1-sol. For gpt-6-sol / gpt-6-luna, adjust it only with reasoning effort set to none; omit it at other effort levels. Omitting effort selects medium, not none.

Number of candidate tokens returned at each position, ranging from 0 to 20; must be used together with include: ["message.output_text.logprobs"].

GPT-6: Astra and 6.1 Sol do not support output logprobs. For Sol / Luna, use them only with reasoning effort none. At other levels, remove logprobs, top_logprobs, and message.output_text.logprobs from Responses include.

Rules for existing models below exclude GPT-6 Sol / Luna and GPT-6.1 Sol:

Note Supported only by the gpt-5.6 family and gpt-5.5; other models do not support this parameter.

Frequency penalty, ranging from -2 to 2, reducing the likelihood of repeated content.

GPT-6 Sol / Luna and GPT-6.1 Sol: support for this parameter has not been confirmed; omit it from basic requests.

Rules for existing models below exclude GPT-6 Sol / Luna and GPT-6.1 Sol:

Note Adjustable on the gpt-5.6 family; the other existing models do not support this parameter. GPT-6 Astra does not allow adjustment and accepts only the default value 0; passing any other value returns 400.

Presence penalty, ranging from -2 to 2, encouraging the model to talk about new topics.

GPT-6 Sol / Luna and GPT-6.1 Sol: support for this parameter has not been confirmed; omit it from basic requests.

Rules for existing models below exclude GPT-6 Sol / Luna and GPT-6.1 Sol:

Note Adjustable on the gpt-5.6 family; the other existing models do not support this parameter. GPT-6 Astra does not allow adjustment and accepts only the default value 0; passing any other value returns 400.

How to handle context that exceeds the window: disabled (default, returns an error) or auto (automatically truncates the middle of the context).

Cache grouping key. GPT-6 / GPT-5.6 route cache traffic automatically; this field is not needed to optimize routing. Separate keys can distinguish cache reuse and accounting for customers or users. Keep the key stable for requests that should reuse a prefix. Earlier models can use stable keys to help cache routing.

Legacy model cache retention. For GPT-6 / GPT-5.6, use prompt_cache_options.ttl: "30m"; do not use 24h in the new field.

Stable identifier of the end user, used for abuse tracking.

GPT-6 Sol / Luna and GPT-6.1 Sol: support for this parameter has not been confirmed; omit it from basic requests.

Rules for existing models below exclude GPT-6 Sol / Luna and GPT-6.1 Sol:

Note Supported only by gpt-6-astra and the gpt-5.6 family; other models do not support this parameter.

End-user identifier, used to distinguish the source of calls.

Unique ID of this response, which can be used as previous_response_id for the next turn

Response type

Response status: completed for a normal ending, incomplete when generation stopped early for reasons such as reaching max_output_tokens, failed when generation failed

Actual model name used

Creation timestamp