I’ve been experimenting with the new reasoning configuration fields in the Responses API and noticed something interesting.
When I set reasoning.effort to "minimal"andreasoning.summary to "auto" or "detailed", I don’t seem to receive any reasoning summary in the output.
Example request
{ “model”: “gpt-5”, “input”: [ { “type”: “message”, “role”: “user”, “content”: [ { “type”: “input_text”, “text”: “Solve this logic puzzle: If all roses are flowers and some flowers are red, can we conclude that some roses are red? Explain your reasoning.” } ] } ], “reasoning”: { “effort”: “minimal”, “summary”: “auto” }, “stream”: false, “store”: false }
My question is:
Does setting effort to "minimal" effectively disable reasoning summaries, even when summary is set to "auto" or "detailed"?
Or is this an unintended behavior / current limitation?
Would love to understand how effort interacts with summary — specifically whether a minimal effort run intentionally omits summaries to save compute or if summaries should still appear regardless of effort level.
Whether you actually receive a reasoning? It would need sufficient unseen reasoning to be done. Minimal is requesting no reasoning, everything short of you getting a non-reasoning AI like gpt-5-chat-latest (a copy of ChatGPT’s model that doesn’t accept reasoning parameters at all).
Summaries should be dependent how much “thinking” there actually is to be reported on, and what strategy you use beyond the input parameter (which is merely a prompt now out of your control) to further minimize how much reasoning is done.
The summary has a style that there is an observer tasked with simply writing a sanitized output, that would need enough “thinking channel” output as its input to be run against to deliver the thinking summary product.
(angry: I notice that gpt-5 in ‘minimal’ is degraded and not delivering partial outputs that have been paid for and requested. I want truncation at 640 tokens by parameter, not the entire thing blocked and billed as “reasoning” by the endpoint.)
input tokens: 105
output tokens: 640
uncached: 105
non-reasoning: 0
cached: 0
reasoning: 640
But I can get zero output billed as reasoning when there is success.
Using targeted developer messaging along with the “minimal” effort reasoning parameter for high assurance that the AI isn’t doing internal deliberations to an output channel other than final:
input tokens: 105
output tokens: 1216
uncached: 105
non-reasoning: 1216
cached: 0
reasoning: 0
With the model almost immediately transitioning to seen output, there is no chance for triggering whatever method is being used to deliver summaries, such as a periodically-triggered rewriter of the output. So it is somewhat expected there is nothing to summarize.
