OpenAI released the Decisions API in beta on October 6, 2026, using GPT-6 Luna to turn text and images into typed decisions. The endpoint is designed for finite answers such as a category, a condition estimate or an ordered score. The API changelog dates the beta milestone.
A smaller output contract can simplify an application
The Decisions guide specifies POST /v1/decisions and three question types: predicate, choice and score. GPT-6 Luna is the currently supported model. OpenAI advertises a speed advantage over Responses; that is a vendor claim to evaluate within the actual application.
A finite output avoids asking a general-purpose assistant to write prose and then parsing a category out of it. That can make routing and prioritization easier to integrate. It also narrows the task: an application needing a detailed explanation, arbitrary structured extraction or tool-call arguments still has a different output requirement.
Probabilities are useful only in context
A returned probability is a model estimate, not proof that an event occurred. Its usefulness depends on the input distribution and how the question is written. A high score on an ambiguous request can still be wrong if the system’s options leave out the correct category.
Evaluate representative examples with known outcomes, including rare but costly errors. Compare thresholds by the consequences of false positives and false negatives. For a support router, sending a request to the wrong queue may be recoverable; for a payment decision, a similar error can have a different cost. Preserve an escalation path for inputs that the defined options do not handle well.
Keep the action boundary outside the classifier
A model can classify a request without being authorized to execute it. The application must independently check the user, resource and permitted action before applying a decision. A label such as cancel or approve should not bypass those checks.
Record the question version and model alongside evaluation results, then test refusals, invalid input and service failure explicitly. Measure complete workflow latency rather than only endpoint response time. The beta is a promising interface for bounded decisions, but its practical value comes from a well-defined decision policy and evidence that it behaves acceptably on the workload.