Decisions API
OpenAI's POST /v1/decisions endpoint, in public beta on gpt-6-luna: it takes text or image evidence plus typed questions and returns probabilities, a choice from a fixed set, or a score instead of prose.
The Decisions API is an OpenAI endpoint, POST /v1/decisions, that returns typed answers instead of generated text. It was announced at DevDay 2026 and opened in public beta on 6 October 2026. At the time of writing gpt-6-luna is the only supported model and OpenAI expects general availability in the coming weeks.
How it works
A request has three parts: the model, the input (text, or messages with text and images; images must be inline base64) and a list of questions. Each question has a type:
- Predicate: the probability, from 0 to 1, that a condition is true.
- Choice: one option from a fixed set you supply, with a probability for each option and a
confidencefield. - Score: a rating against ordered levels, returned as the probability-weighted average of the level indices, with probabilities and
confidence.
Any answer can come back as a refusal. Use it to classify content, route requests or choose an agent's next action.
Price and limits
Input costs $0.10 per million tokens; output and cache tokens are not billed, long inputs above 272K tokens cost double and regional processing adds 10%. OpenAI describes the endpoint as about 10x faster than the Responses API but publishes no benchmark, and it publishes no accuracy or calibration figures. Treat the returned numbers as inputs to your own threshold policy and check them with calibration on your own data.
Related terms: decision layer, confidence score.
For the strategic argument about who owns the threshold, see OpenAI Decisions API turns probability into policy.
Related terms
Decision layer
A dedicated component between AI judgment and organizational action that decides whether the work should continue, separate from the part of the system that does the work.
Threshold policy
The explicit rules that turn a model's probability or score into an action — reject, send to a human, or act autonomously — owned and reviewed separately from the model itself.
Calibration
How well a model's stated confidence matches reality: if it says 90% on a hundred comparable cases, roughly ninety should be correct. Once probabilities drive actions, it becomes an operating metric.
Confidence score
A number a model or classifier reports to indicate how sure it is. It is only useful for decisions if it is calibrated, and it never decides on its own what action is permitted.