• Sources: primary, discussion
  • Summary: The endpoint returns typed answers rather than generated text, in three question forms: predicate returns a probability from 0 to 1, choice selects from supplied values, and score returns a probability-weighted average over ordered levels. OpenAI states it is about 10 times faster than the Responses API and that gpt-6-luna is the only model currently available. It is a public beta with general availability stated as weeks away, so it is not yet a stable dependency.
  • Why it matters: This changes the cost of classification, routing and prioritization steps that today burn a full generation call, and the numeric probability on a predicate is the part structured outputs does not give.