I remember a couple weeks ago, the completion responses from o1 and similar reasoning models used to include “reasoning_content” in the completion responses but those are no longer available? Were they removed recently, or am I remembering wrong, and they were never available to begin with?
Searching in this forum I’m surprised how few people (nobody?) is even curious about this 
I think you are remembering wrong. 
When I accessed the o3-mini model via API today, I got the reasoning_content field in the response, but the content was empty.
I guess OpenAI has temporarily closed (or has not had time to open?) the ability to get reasoning content through the API
The response is like:
ChatCompletionMessage(content='', refusal=None, role='assistant', audio=None, function_call=None, tool_calls=[ChatCompletionMessageToolCall(id='call_alzIzla0JrnpqvicBQVYCKnR', function=Function(arguments='{"MY_TOOL": 123}', name='MY_TOOL_NAME'), type='function')], reasoning_content='')
you can see there is reasoning_content field at the end of message.
Hi there, I’m using the Strands Pyton SDK to build agents and it allows us to capture the ‘reasoning_content’ from LLMs but I can’t find a sample from OpenAI that returns this.
Does anyone know which model returns this in the response? Have tried reasoning models as well but don’t see it in the response.
You need to:
1: read documentation
https://platform.openai.com/docs/guides/reasoning#reasoning-summaries
2: have subjected yourself to the intrusion of personal ID verification, not “for safe deployment”, but rather “personal information waiting for security breach” that no security professional would advise.
3: Use responses endpoint
4: Use o3 or o4-mini
5: Use correct tool parameters
6: Have code that receives output list of objects and parses the types.
1 Like
Thank you for your response! I didn’t quite understand #2 but everything else is clear.
The tool I am using ‘Strands’, specifically expects a reasoningContent in the LLM response, so I may have to intercept the response to add the response summary where the tool expects it.