OpenAI answers TypeSafe’s Jev with a Decision API built on Luna

OpenAI has unveiled its own Decision API, a model that leverages the company's Luna model, following the surge in popularity of TypeSafe's Jev decision models. The announcement was made at OpenAI's DevDay conference. As a preview, the Decision API is currently available for limited testing, with a broader rollout expected in the coming days.

The Decision API is designed to return a set of predefined answers with confidence scores, rather than generating text like traditional language models. This makes it particularly well-suited for tasks like content classification, request routing, or choosing an agent's next course of action from a limited set of options. The model is reported to return results in just 150 milliseconds, significantly faster than OpenAI's GPT-6 Luna model, which takes 1.6 seconds.

To use the Decision API, developers only need to provide a set of questions, answers, and context. In contrast, traditional approaches often involve using a chat model with a prompt that instructs it to select the correct answer from a list, or training a separate small classifier, which requires labeled data and frequent retraining when the label set changes.

The Decision API represents a middle ground between these two options. It allows developers to incorporate new labels into the prompt and receive a confidence score, which they can then use in their applications. While OpenAI has previously offered a Moderation API that returns per-category scores, this new API allows developers to set their own categories.

However, there are still some key details that remain unclear. The cost per API call, the maximum number of candidate answers per request, and the ability for developers to fine-tune the model on their own data are all factors that will ultimately determine whether the Decision API becomes a standard component in agent frameworks or remains a niche tool for specialized use cases.