response.reasoning_summary_text.delta incorrectly uses ResponseAudioDeltaEvent

[simonw]

Confirm this is an issue with the Python library and not an underlying OpenAI API

• This is an issue with the Python library

Describe the bug

In the streaming output the response.reasoning_summary_text.delta event is represented using ResponseAudioDeltaEvent for some reason.

To Reproduce

Try running this (STR created with the help of o4-mini):

python -c 'import openai,os
for event in openai.OpenAI(api_key=os.getenv("OPENAI_API_KEY")).responses.create(
    model="o3",
    input=[{"role":"user","content":"Why is the sky blue?"}],
    reasoning={"summary":"auto"},
    stream=True,
):
    print(event)'

The output includes lines like this:

ResponseAudioDeltaEvent(delta='**Expl', type='response.reasoning_summary_text.delta', item_id='rs_680044dcbd9881918572b2b1a63ca6a903b0150bf71ce6fa', output_index=0, summary_index=0)

ResponseAudioDeltaEvent looks wrong. Something like ResponseReasoningSummaryDeltaEvent would make more sense.

OS

macOS

Python version

Python 3.13.2

Library version

openai==1.75.0

[lemeb]

Can confirm on my end! Very weird.

[lemeb]

Ok, I think I figured what's going on:

Because events of type response.reasoning_summary_text.delta don't have any data structure associated with it, the discriminator mapping fails.

This line of code:

openai-python/src/openai/_models.py

Line 497 in ed53107

discriminator = _build_discriminated_union_meta(union=type_, meta_annotations=meta)

Will yield this:

Click to see the value of `disciminator`

{'error': <class 'openai.types.responses.response_error_event.ResponseErrorEvent'>,
 'response.audio.delta': <class 'openai.types.responses.response_audio_delta_event.ResponseAudioDeltaEvent'>,
 'response.audio.done': <class 'openai.types.responses.response_audio_done_event.ResponseAudioDoneEvent'>,
 'response.audio.transcript.delta': <class 'openai.types.responses.response_audio_transcript_delta_event.ResponseAudioTranscriptDeltaEvent'>,
 'response.audio.transcript.done': <class 'openai.types.responses.response_audio_transcript_done_event.ResponseAudioTranscriptDoneEvent'>,
 'response.code_interpreter_call.code.delta': <class 'openai.types.responses.response_code_interpreter_call_code_delta_event.ResponseCodeInterpreterCallCodeDeltaEvent'>,
 'response.code_interpreter_call.code.done': <class 'openai.types.responses.response_code_interpreter_call_code_done_event.ResponseCodeInterpreterCallCodeDoneEvent'>,
 'response.code_interpreter_call.completed': <class 'openai.types.responses.response_code_interpreter_call_completed_event.ResponseCodeInterpreterCallCompletedEvent'>,
 'response.code_interpreter_call.in_progress': <class 'openai.types.responses.response_code_interpreter_call_in_progress_event.ResponseCodeInterpreterCallInProgressEvent'>,
 'response.code_interpreter_call.interpreting': <class 'openai.types.responses.response_code_interpreter_call_interpreting_event.ResponseCodeInterpreterCallInterpretingEvent'>,
 'response.completed': <class 'openai.types.responses.response_completed_event.ResponseCompletedEvent'>,
 'response.content_part.added': <class 'openai.types.responses.response_content_part_added_event.ResponseContentPartAddedEvent'>,
 'response.content_part.done': <class 'openai.types.responses.response_content_part_done_event.ResponseContentPartDoneEvent'>,
 'response.created': <class 'openai.types.responses.response_created_event.ResponseCreatedEvent'>,
 'response.failed': <class 'openai.types.responses.response_failed_event.ResponseFailedEvent'>,
 'response.file_search_call.completed': <class 'openai.types.responses.response_file_search_call_completed_event.ResponseFileSearchCallCompletedEvent'>,
 'response.file_search_call.in_progress': <class 'openai.types.responses.response_file_search_call_in_progress_event.ResponseFileSearchCallInProgressEvent'>,
 'response.file_search_call.searching': <class 'openai.types.responses.response_file_search_call_searching_event.ResponseFileSearchCallSearchingEvent'>,
 'response.function_call_arguments.delta': <class 'openai.types.responses.response_function_call_arguments_delta_event.ResponseFunctionCallArgumentsDeltaEvent'>,
 'response.function_call_arguments.done': <class 'openai.types.responses.response_function_call_arguments_done_event.ResponseFunctionCallArgumentsDoneEvent'>,
 'response.in_progress': <class 'openai.types.responses.response_in_progress_event.ResponseInProgressEvent'>,
 'response.incomplete': <class 'openai.types.responses.response_incomplete_event.ResponseIncompleteEvent'>,
 'response.output_item.added': <class 'openai.types.responses.response_output_item_added_event.ResponseOutputItemAddedEvent'>,
 'response.output_item.done': <class 'openai.types.responses.response_output_item_done_event.ResponseOutputItemDoneEvent'>,
 'response.output_text.annotation.added': <class 'openai.types.responses.response_text_annotation_delta_event.ResponseTextAnnotationDeltaEvent'>,
 'response.output_text.delta': <class 'openai.types.responses.response_text_delta_event.ResponseTextDeltaEvent'>,
 'response.output_text.done': <class 'openai.types.responses.response_text_done_event.ResponseTextDoneEvent'>,
 'response.refusal.delta': <class 'openai.types.responses.response_refusal_delta_event.ResponseRefusalDeltaEvent'>,
 'response.refusal.done': <class 'openai.types.responses.response_refusal_done_event.ResponseRefusalDoneEvent'>,
 'response.web_search_call.completed': <class 'openai.types.responses.response_web_search_call_completed_event.ResponseWebSearchCallCompletedEvent'>,
 'response.web_search_call.in_progress': <class 'openai.types.responses.response_web_search_call_in_progress_event.ResponseWebSearchCallInProgressEvent'>,
 'response.web_search_call.searching': <class 'openai.types.responses.response_web_search_call_searching_event.ResponseWebSearchCallSearchingEvent'>}

And so this code will not return anything:

openai-python/src/openai/_models.py

Lines 498 to 503 in ed53107

	if discriminator and is_mapping(value):
	variant_value = value.get(discriminator.field_alias_from or discriminator.field_name)
	if variant_value and isinstance(variant_value, str):
	variant_type = discriminator.mapping.get(variant_value)
	if variant_type:
	return construct_type(type_=variant_type, value=value)

Which means you rely on this fallback code, which essentially ends up selecting the first element in the mapping (ResponseAudioDeltaEvent), and coerce a construction.

openai-python/src/openai/_models.py

Lines 505 to 511 in ed53107

	# if the data is not valid, use the first variant that doesn't fail while deserializing
	for variant in args:
	try:
	return construct_type(value=value, type_=variant)
	except Exception:
	continue

[lemeb]

ok, seems like it was fixed by 6321004!

[moonbox3]

This is happening again in #2382 and #2383

[apurvsibal]

+1

[abhinav-wingback]

+1