OpenAI has introduced a new Decisions API that uses its Luna model to choose between predefined answers for tasks such as classification, routing, and agent actions. The limited-preview tool is designed to make these decisions faster and cheaper than using a general-purpose model to generate a full response.
The company revealed the API during DevDay 2026. CEO Sam Altman said narrowing the model to a defined choice allows OpenAI to increase speed while preserving capabilities such as image understanding, multilingual support, and safety protections.
Decisions API Targets Structured Software Tasks
Rather than asking a model to generate open-ended text, developers can provide a set of possible answers and have Luna select among them. OpenAI says the API can be used to classify inputs, route requests, or decide which action an AI agent should take.
The approach is similar to Jev, a recently released model from TypeSafe AI that is built specifically for software automation. TypeSafe describes Jev as a “System One Model” that takes unstructured inputs and returns predefined, typed decisions alongside probability and confidence scores.
TypeSafe says Jev avoids generating open-ended strings and instead produces all of its outputs in parallel. The company claims that design makes it substantially faster and cheaper than conventional large language models for some automation workloads, although those performance claims come from TypeSafe’s own evaluations.
Developers Are Testing Decision Models for Agent Safety
TypeSafe founder Diogo Almeida, a former OpenAI researcher, has argued that these faster decision models could supplement larger models rather than replace them. Developers are already testing Jev for tasks such as command classification, workflow routing, and monitoring AI agents.
One example comes from QueryStory founder Shapor Naghibzadeh, who built a demo that checks individual agent actions against the task the agent was assigned. The system can permit actions, flag uncertain ones for review, or block those it judges to be inappropriate.
In that demonstration, monitoring cost $2.94 using Jev compared with $372 using a frontier LLM, according to the developer. The result suggests that lower-cost decision models could make it practical to evaluate every action an AI agent takes, though the comparison comes from a prototype rather than independent production testing.
OpenAI has released Decisions API only as a limited preview, so independent comparisons with Jev remain scarce. The company has not yet published detailed public benchmarks showing how accurately its probability outputs correspond with real-world outcomes.
Featured image credits: Wikimedia Commons
For more stories like it, click the +Follow button at the top of this page to follow us.