Will OpenAI add logprobs support to the Chat Completions API for GPT-3.5 and GPT-4?

Below is a summary of the current situation:

Current Support

• Logprobs in the Completions Endpoint:
  The logprobs option is available in the legacy Completions API (e.g., with the davinci-family models). This parameter lets you obtain log probabilities for each token, which is useful for scoring completions.

• Chat Completions Endpoint:
  For the gpt-3.5 and gpt-4 model families—which are only accessible via the Chat Completions API—the logprobs parameter is not available at this time.

Future Plans

At the moment, no public announcements have been made regarding plans to support log sampling probabilities via the Chat Completions API. OpenAI continues to evaluate and iterate on its offerings, so while it’s possible that similar functionality could be introduced in the future, there’s no confirmed timeline or commitment to do so.

Recommended Approach

If scoring completions is a crucial part of your workflow, you might consider:

• Using the Legacy Completions API:
  Continue using models that support the logprobs parameter.
• Monitoring Updates:
  Keep an eye on the official changelog and API documentation (e.g., OpenAI’s changelog) for any feature updates or changes.

Feel free to ask more questions about the API if needed!