OpenAI introduced the Decisions API at its DevDay 2026 event on September 29. The tool is designed to provide quick, constrained responses based on a list of options defined by developers, rather than generating long-form text.
The API is powered by a specialized version of the GPT-6 Luna model. It processes both text and image inputs, allowing developers to categorize data or determine the next action for AI agents by selecting from a predefined set of choices.
Early demonstrations indicate a significant speed advantage, with response times of approximately 150 milliseconds compared to 1.6 seconds for standard Luna calls. This represents a latency improvement of roughly 10 times.
The API is currently available in a limited preview for select customers, with a broader release expected in the coming days. OpenAI has not yet disclosed pricing, accuracy benchmarks, confidence scores, or specific request formats.
The launch aligns with the company's focus on improving AI agent infrastructure. By handling routing and classification tasks, the Decisions API aims to enable more efficient architectures where lightweight models manage traffic between complex reasoning steps.
