OpenAI launches decision-classification API to oversee and secure AI agents

During its Dev Day event this week, OpenAI introduced a Decisions API that sits alongside the Luna model to enable rapid classification and decision-making at a fraction of the cost of traditional large language models. The tool is designed to offer the Luna model a fixed set of choices to evaluate—say, image categories or different agent behaviors—and return probabilities rather than full text generation.

CEO Sam Altman framed the capability as a way to "make it extremely fast while keeping capabilities like image understanding, broad language support, and safety protections." The announcement drew immediate parallels to Jev, a similar decision-classification model released by TypeSafe AI just weeks ago.

What it is#

Jev and similar tools represent a shift in how developers think about AI architecture. Rather than asking a general-purpose language model to pick from options—a wasteful operation that generates tokens unnecessarily—these decision models are explicitly trained to evaluate bounded choice sets quickly and cheaply. TypeSafe CEO Diogo Almeida, a former OpenAI researcher, quipped on social media about the "beginning of the clone wars," but acknowledged that OpenAI's move signals industry recognition that "building in a System One compatible way is the future."

The terminology draws from cognitive science: System One models handle fast, intuitive classification, while System Two handles deliberate reasoning. For many software automation tasks, System One suffices and costs far less.

The cost advantage#

Traditional frontier LLMs are comparatively expensive and slow for simple classification tasks. Developers working with Jev have found substantial savings. One concept gaining traction is using such decision models as a safety layer for AI agents: monitoring each action the agent takes against its original task, blocking suspicious moves, and flagging others for human review.

A demo built for a recent hackathon showed that same kind of monitoring on a Hugging Face scenario would cost $2.94 with Jev versus $372 with a frontier LLM—a more than hundredfold reduction. At such a price point, it becomes feasible to inspect every single agent action, providing a blanket security layer that could prevent the kind of agent misbehavior incidents OpenAI has faced.