Logprobs deprecated for gpt-5 models?

API Deprecations
1

Hi all,

We were previously logprobs as an indicator of RAG answer confidence. However, since switching to gpt-5-mini from 4o/4o-mini we have had to remove the logprobs=True from our payload, since this appears to be fully deprecated from the API…

Is there any intention to bring logprobs back in the near future, or alternatively, does anyone have any other methods for confidence scoring?

1 Like
2
  • Logprobs have not been available for any “reasoning” models, a symptom or choice seen since the introduction of o1-preview.
  • gpt-5 is a reasoning model, just without the “o” scheme, and not really fitting into the progressing performance path up to o4-mini.
  • There is a variant gpt-5-chat that has more typical model parameters by not actually reasoning, but it also does not deliver logprobs. It is likely the concept was discarded, because OpenAI has been already been running a completely parallel “fake” softmax to deliver logprobs that hide information about special tokens and mask probabilities for tokens that were sampled from beneath the top-20.