OpenAI Responses API driver for Responses endpoint models.

Implements the ReqLLM.Providers.OpenAI.API behaviour for OpenAI's Responses endpoint, which provides extended reasoning capabilities for advanced models.

Endpoint

/v1/responses

Supported Models

Models with "api": "responses" metadata:

• o-series: o1, o3, o4, o1-preview, o1-mini
• GPT-4.1 series: gpt-4.1, gpt-4.1-mini
• GPT-5 series: gpt-5, gpt-5-preview

Capabilities

• Reasoning: Extended thinking with explicit reasoning token tracking when supported
• Streaming: SSE-based streaming with reasoning deltas and usage events
• Tools: Function calling with responses-specific format
• Reasoning effort: Control computation intensity (minimal, low, medium, high)
• Enhanced usage: Separate tracking of reasoning vs output tokens

Encoding Specifics

• Input messages use input_text content type instead of text
• Token limits use max_output_tokens instead of max_tokens
• Tool choice format: {type: "function", name: "tool_name"}
• Reasoning effort: {effort: "high"} format
• Code Interpreter: tool maps such as %{"type" => "code_interpreter", "container" => ...} are passed through unchanged to OpenAI. Both object containers (%{"type" => "auto", ...}) and string container IDs are supported. This is Responses-API-only; Chat Completions does not support the code_interpreter tool type.

Code Interpreter

Raw code_interpreter_* output items from the response are collected in response.provider_meta["code_interpreter"]["items"] and are excluded from normal text and function tool-call extraction.

Decoding

Non-streaming Responses

Aggregates multiple output segment types:

• output_text segments → text content
• reasoning segments (summary + content) → thinking content
• function_call segments → tool_call parts
• refusal message parts → text content, finish_reason: :content_filter, and the refusal texts in provider_meta["refusals"] (see ReqLLM.Response.refusals/1)

Streaming Events

• response.output_text.delta → text chunks
• response.refusal.delta → text chunks; response.refusal.done emits nothing
• response.reasoning.delta / response.reasoning_text.delta → thinking chunks
• response.reasoning_summary_text.delta → thinking chunks whose metadata carries item_id, output_index and summary_index
• response.reasoning_summary_part.added / .done → meta chunks with a reasoning_summary_part map (status, item_id, output_index, summary_index, text) marking summary part boundaries
• response.output_item.done with a compaction item → :content_part chunk holding a :provider_block that later requests replay verbatim
• response.usage → usage metrics with reasoning_tokens
• response.completed → terminal event with finish_reason, :content_filter when the output holds a refusal, whose texts are also put in provider_meta["refusals"]
• response.incomplete → terminal event for truncated responses
• response.failed → terminal event with finish_reason: :error and the failure message
• error → terminal event with finish_reason: :error and the error message

Usage Normalization

Extracts reasoning tokens from usage.output_tokens_details.reasoning_tokens and provides:

• :reasoning_tokens - Primary field (recommended)
• :reasoning - Backward-compatibility alias (deprecated)