OpenAI introduces Decisions API for structured software automation

At a developer event on Tuesday, OpenAI previewed a new capability called the Decisions API. Designed to constrain model outputs to a predefined set of choices, the API aims to balance high-speed execution with core language and image understanding.

Understanding the approach

The interface functions similarly to existing classification tools on the market, such as Jev, which was released by TypeSafe AI earlier this month. These systems are built to output probabilities quickly and cheaply, offering an alternative to traditional, slower large language models for specific software automation tasks.

During the announcement, leadership noted that the API allows models like Luna to select from designated options, ranging from image categories to specific agent behaviors. By narrowing the scope of the model's task, developers can achieve significantly faster response times while retaining necessary safety protections.

Practical applications for agents

A primary use case for these streamlined classification models is the monitoring and security of autonomous AI agents. Recent incidents involving misbehaving agents on the open internet have highlighted the need for robust oversight. Traditionally, this supervision requires a separate frontier model running at a substantial compute cost.

Industry observers note that lightweight decision models could reduce these monitoring expenses considerably. In experimental settings, running a fast classifier to check agent actions against assigned tasks costs a fraction of operating a full-scale LLM for the same verification process.

This cost efficiency makes continuous oversight feasible for nearly every agentic action. As other startups introduce similar classification tools, the focus across the industry is shifting toward optimizing the intelligence-per-dollar curve, ensuring that fast automation does not compromise accuracy or safety.