O3 model in API often omits reasoning summary despite reasoning.summary: detailed

API
,
21

Same here, and the support is not helping…

22

Ack, sorry to hear you haven't been able to get help from Support - happy to try debugging here. Would you be able to share any recent request_ids + timestamps from calls that are missing reasoning.summary?

23

An example from GPT-5 with one empty reasoning summary:

resp_6897cb436d048197b949a38ace6187050082566ebe65d76a - Aug 9, 2025, 6:27 PM EDT

An earlier example from o3 with literally 12 empty reasoning summaries:

resp_6891ec07251081989379f6197396dc0a0a8eae07db3f7d2e - Aug 5, 2025, 7:33 AM EDT

(Presume those are safe to share!)

Chris

24

Hi Chris,

Thank you! Not finding anything with those IDs just yet.

Would you mind writing into support@openai.com with the email that's attached to your forum account and letting me know here once you've done that? That way, I can personally track down your message and reply directly.

25

Just did! And sorry, yes, I had emailed support under a separate organizational email address than this forum account! Just emailed again under this email…

Thanks!

Chris

26

No worries, thank you! Will follow up with you there and we can post our learnings here if/when we have any.

27

I am getting the same issue of empty reasoning summaries prior to tool calls on gpt-5. I can see a reasoning item was created. But the summary is empty. I have reasoning={"summary": "auto"}.

28

After more testing, I found that the o3 model’s reasoning summaries would be broken by the web_search_preview tool. For gpt-5, I wouldn’t get any reasoning summaries by default, but if I specified “auto” then I would get at least 50% or so of them (still many blank, but many non-blank as well).

29

Follow up question- what is the purpose of seeing the reasoning summaries? I always turn them off.

30

For my application, it’s really important that there be transparency in both sources and steps, to build trust with users.

Chris

31

Thanks chris, if you have time i’d be curious to hear if it is even feasible to read through the reasoning in responses with huge outputs. I iterate a lot (massive amounts of tokens per second), and if I were to also consider the “reasoning steps & sources” I would never get anything done. What are your thoughts?

32

Oh, this isn’t for me, it’s for my product’s users. And while I don’t expect every user to carefully inspect sources and reasoning for every job they run through my agent. But being transparent in what happened and why (to the extent the curtain can be pulled back on the “why” part) is generally important for my users — particularly the more AI-skeptical ones.

Chris

33

I’m not getting any summaries on gpt-5 with reasoning set to auto. Basically all interactions with my agent lead to a tool call. I wonder if the reasoning pre-tool call is less stable as a lot of people noted with o3.

34

No summaries for me either. Not great that this topic is marked as resolved when it clearly isn’t…

When using the Responses API

The following events are rarely emitted on o3

  • response.reasoning_summary_part.added
  • response.reasoning_summary_part.done
  • response.reasoning_summary_text.delta
  • response.reasoning_summary_text.done

And the following events are NEVER emitted in my experience

  • response.reasoning_text.delta
  • response.reasoning_text.done
35

I’m having the same issue. I also tried with o4-mini but nothing changed.

36

Same for me.

Only working consistently with GPT-5.

Not working for me for:

  • o3
  • o3-pro

Doing same like with GPT-5 in the request.

        reasoning: {
          effort: 'high',
          summary: 'auto',
        }
...
37

The OpenAI support continue to send me an automated response and it not helps to resolve the problem…

38

Hi everyone!

Catching up here. Here’s where things stand internally:

  • The behavior you’re running into ( o3 or o3-pro sometimes omits reasoning.summary) is known and expected under certain request conditions, particularly when tools is included.

  • We have rolled out an improved reasoning summary pipeline for GPT-5 models, and we plan to back-port those improvements to o3 and o3-pro.

  • In practice, leaving out the tools parameter entirely (rather than sending an empty or null list) tends to produce reasoning summaries more reliably.

  • You must also have your organization verified in the API settings for reasoning summaries to work.

  • That said, even with optimal setup, there may still be occasional omissions until the full fix is live.

If you want more consistent reasoning output today, you might try using gpt-5-pro, which already incorporates the updated summarization logic.

Apologies in advance if I’ve missed any context in this thread and please do let me know if there’s anything else I can try to clarify. Thanks for all the discussion here so far.

39

Hello @vc-openai,
Is this issue fully addressed in gpt-5.2? Facing the same issue, sometimes the reasoning summary is empty.

40

Hi! GPT-5.2 is part of the GPT-5 family that already received an improved reasoning-summary pipeline, so it will usually be more reliable than older models. However, it seems there are occasional (and expected) empty summaries in some request shapes, particularly when a tools parameter is present (or when an empty/null tools list is sent).


A few things to try and check:

  • Omit the tools parameter entirely rather than sending tools: []
  • Confirm your organization is verified in the API settings (reasoning summaries for GPT-5 models require verification)
  • If you need more consistent reasoning summaries today, try gpt-5-pro or gpt-5.2 with an appropriate reasoning_effort setting

If you’re still seeing empty summaries, could you share the exact model you’re calling, whether you’re sending tools (and what value), your reasoning.* settings, and a minimal reproducible request (request IDs / timestamps help)? That will let us look for the request in logs and triage further. Thank you!