OpenAI Launches Decisions API for Cheap Agent Monitoring

OpenAI used its DevDay event to quietly reveal a new "Decisions API," a lightweight classifier designed to make agent workflows faster and cheaper to monitor. The move closely mirrors Jev, a fast reasoning model from startup TypeSafe AI that is already winning over developers who need to screen agent actions at scale.

What happened

During DevDay, CEO Sam Altman introduced the Decisions API as a way to hand OpenAI's Luna model a predefined set of choices and have it return probabilities quickly and at low cost. Rather than running a full frontier model for every decision, the API focuses the model on a single choice, which Altman said keeps capabilities like image understanding and safety protections while cutting latency and cost.

The product strongly resembles Jev, a model TypeSafe AI shipped earlier this month for software automation. Developers use Jev to augment larger LLMs for tasks where a full model is overkill, and they report it is both faster and cheaper. TypeSafe CEO Diogo Almeida, a former OpenAI engineer, joked on X about "the beginning of the clone wars," adding that OpenAI's interest signals "building in a System One compatible way is the future."

Why cheap classifiers matter

The most compelling application is watching over AI agents. After a series of incidents in which its agents misbehaved on the open internet, OpenAI began using a separate model to review agent actions for bad behavior, at what the company called "significant compute cost."

A cheap classifier changes that math. Security researchers estimate that screening every agentic action against its assigned task with Jev costs about $2.94, versus roughly $372 using a frontier LLM. At that price, running a reviewer on every single action becomes practical, not just the important ones.

OpenAI released the Decisions API only as a limited preview, and it remains unclear how closely it will match Jev's capabilities. But the direction is unmistakable: the agent stack is splitting into fast, cheap classifiers for high-frequency decisions and large reasoning models for the hard ones.

What this means for builders

For teams building and operating AI agents, the takeaway is architectural. Not every step in an agent loop needs a frontier model. A dedicated classifier that returns probabilities in milliseconds can gate actions, flag anomalies, and enforce policy at a fraction of the cost. According to TechCrunch, the Decisions API is already drawing interest from developers, and the pattern is worth evaluating in any production agent workflow.