OpenAI has opened a new public beta that turns a common prompt-engineering chore into a typed API call.
The Decisions API, announced on 6 October, exposes a single endpoint — POST /v1/decisions — that takes text or images and returns structured answers in three shapes. A predicate estimates, as a probability between 0 and 1, whether a stated condition holds. A choice picks one value from a set the developer supplies, with a confidence attached to each. A score returns a number, for things such as severity, priority or damage. Instead of writing a prompt and parsing free-form prose back out, developers describe the decision they want and get a value they can branch on directly.
The intended jobs are unglamorous and everywhere: classifying user content, routing a support request to the right queue, triaging complaints, flagging a photo for damage, or scoring how urgent a ticket is. Because a low-confidence answer can be escalated to a stronger model, the cheap path can handle the easy majority of cases and hand the ambiguous rest upwards — exactly the tiering that API vendors want applications to be built on.
Early accounts say the beta runs on a single model, reported as gpt-6-luna, rather than the model-picker pattern of some earlier OpenAI tooling. Whatever the backing, the design point is the same: make the vast middle of AI workloads — the small, repeated judgements that never needed a conversation — cheap, fast and predictable. Reports circulating in developer forums claim the endpoint is substantially faster than coaxing the same answer out of a chat completion, though such figures come from the vendor and early testers rather than independent measurement.
The Decisions API is in public beta and available to developers now.
