In the DevDay keynote, Sam Altman spent only a few sentences introducing the Decisions API: it lets the model respond in a fraction of a second, and it works by giving the Luna model a set of preset options and pushing response speed to the extreme by making the model focus on that single choice.
This passage contains two key pieces of information. First, the Decisions API narrows open-ended generation into "choosing from explicit options," which is highly consistent in design philosophy with Jev's Choice primitive. Second, it chose Luna as the underlying model — the smallest and cheapest in OpenAI's current model lineup. Using the smallest, lowest-cost model to handle decision tasks rather than calling a larger model shows that this capability is clearly aimed at the scenario of a high-frequency, low-cost foundational component that can be embedded into Agent workflows at scale.
But this new API is currently available only as a limited preview, and pricing, the number of candidate answers supported per request, and whether developers can tune it with their own data are all undisclosed. An OpenAI spokesperson told foreign media that the company plans to release more information "at full launch." It can therefore be judged that OpenAI most likely rushed this launch in order to get the news out before DevDay.
From a product positioning standpoint, this kind of capability fills the gap between ordinary chat models and traditional classifiers. In the past, if developers wanted a model to choose among several answers, they usually called an ordinary chat model and constrained the output with prompts, or read token probabilities; another route was to train a dedicated small classifier, which is fast and cheap but requires labeled data, and often needs retraining whenever the labels change. The Decisions API sits between the two: developers can supply new questions and labels through prompts just as with a large language model, while getting structured results and confidence scores similar to a classifier.