OpenAI opens Decisions API for bounded application decisions

OpenAI announced general availability on October 22, 2026. Developers specify the question and permitted answers, then supply text, images or structured JSON data. That can help an application classify an incoming request and choose its next route. Requests now support up to 400 questions; prompt caching reuses context.

The developer guide separates selecting an action from performing it: the application executes the selected action. Independent questions can share one request, but a decision that depends on an earlier answer needs a separate request. The endpoint does not accept audio or files directly. For routing or review thresholds, OpenAI recommends examples from the application with known labels, considering the costs of false positives and missed cases. These examples help set the boundary between automatic routing and sending a case for review.

The voice integration guide describes Live calling Decisions directly. Delegation setup requires a working Live connection and permission to write Decisions requests. Separate transcript triggers can select an avatar’s expression or gesture as the user speaks or Live replies; they operate independently of delegation. When filtering WebRTC events, developers must allow both the decision event and its corresponding transcript event to receive the trigger result. Results can arrive out of order, so the guide instructs applications to apply the newest relevant result for each direction.