OpenAI's Decisions API Gives Luna a Fast Multiple-Choice Mode
OpenAI CEO Sam Altman announced a new Decisions API at OpenAI's Dev Day. The interface is available as a limited preview. TechCrunch
The API works in a similar way to Jev, a model from TypeSafe AI. Jev is built for software automation and turns developer-provided choices into probabilities, cheaply and at high speed.
Altman described the Decisions API as a way to give OpenAI's Luna model a fixed set of options to choose between. Examples included image-classification categories and different agent behaviors. The setup is discriminative rather than generative, like a multiple-choice test rather than an essay. The developer defines the label set, the model selects from it.
OpenAI's DevDay 2026 Recap states that the Decisions API focuses Luna's intelligence on a specific set of user-defined questions. The same recap states that it enables real-time decision-making. OpenAI
Implementation details point to a lightweight classifier path. The Decisions API is built on the small Luna model and returns preset answers with confidence scores. Responses arrive in 150 milliseconds. The New Stack
A 150 millisecond round trip with a closed label set and calibrated scores can be used in request routing, policy selection, moderation queues, and tool-use loops. Developers can set thresholds, abstain, or escalate without parsing free text.
Following incidents where its agents misbehaved on the open internet, one of OpenAI's new security measures is using a separate model to watch for bad actions at significant compute cost.
The broader context here is the shift from open-ended generation to bounded choice in agentic stacks. Free-form reasoning is strong for planning and synthesis. It is awkward for the hundreds of small branch points that decide what an agent does next. A dedicated decision endpoint separates those jobs, with one model proposing actions and a smaller, faster model constraining selection.
For engineers working with agents, the change worth watching is control. Preset options with confidence scores allow fixed rules around unpredictable models. Teams can log distributions, tune thresholds per category, A/B test behavior sets, and enforce least-privilege action spaces. That is normal practice in classification systems, now applied to agent orchestration.
In my view, the comparison to Jev is instructive rather than incidental. If cheap, fast probabilistic choice becomes a basic primitive, frontier capability matters less at the decision point than calibration, latency, and cost per call. The large model can stay centered on reasoning while the small model governs execution. Over time, that split could make agents more predictable to deploy, which helps wider use in software automation where errors compound quickly.
