OpenAI previews Decisions API for fast, bounded model choices
The limited preview asks OpenAI's Luna model to choose among predefined options, a narrower task than an open-ended chat response.
The task OpenAI is targeting
OpenAI previewed a Decisions API at DevDay on September 29, TechCrunch reports. CEO Sam Altman described giving the company’s Luna model a predefined set of choices rather than asking it for an unrestricted text answer. His examples included choosing categories for an image and selecting among agent behaviors. Altman said the narrower task is intended to produce fast responses while retaining capabilities such as image understanding, language support and safety protections. Those are product claims from the announcement, not independently measured results.
TechCrunch compares the approach with TypeSafe AI’s Jev, which also returns structured decisions for software workflows. The comparison describes a category of use, not proof that OpenAI copied Jev or that the two products perform alike. A bounded choice may be useful where software needs a category, score or action selection repeatedly, including inside an AI agent.
What has not been tested publicly
The API is in limited preview, according to TechCrunch. The publication says it has not yet seen developers test it publicly. That leaves important questions about access, pricing, output calibration and real-world speed unanswered in the reviewed source. TechCrunch discusses agent monitoring as a possible application, but it does not report that Decisions API already powers OpenAI’s security systems or that it would have prevented any past incident.