AI
OpenAI launches Decision API to compete with TypeSafe's classification approach
OpenAI has unveiled a new Decisions API built on its Luna model, designed to classify content and route requests with real confidence scores rather than relying on traditional language models.
The New Stack
OpenAI introduced its Decisions API at DevDay on Tuesday, marking the company's entry into the rapidly growing market for decision-focused models. The new tool is built on Luna, OpenAI's smallest and most cost-effective model, and represents the company's response to growing interest in specialized classification systems like TypeSafe's Jev.
How decision models differ from chat models#
Decision models operate on fundamentally different principles than traditional language models. Rather than generating prose, these systems accept a question, a set of predefined answers, and relevant context—then return one selected answer paired with an actual confidence score.
This approach addresses a key weakness of general-purpose LLMs. Conventional chat models can theoretically be asked to pick from a list, but their confidence scores are rough approximations, and the process consumes substantial token overhead. Decision models instead deliver results in around 150 milliseconds with genuine probability estimates, compared to 1.6 seconds for Luna operating as a standard chat interface.
Ideal use cases for classification and routing#
These models excel at specific tasks where choosing from a fixed set of options matters: content moderation, request routing, and determining an AI agent's next action from a limited menu of choices. They function somewhere between two existing approaches developers have relied on—writing careful prompts for a general chat model, or training a small, specialized classifier that requires labeled training data and retraining each time categories shift.
Current status and open questions#
The Decisions API is available now in limited preview, with broader availability rolling out in coming days. OpenAI has shared few technical details so far, promising more information during the broader release.
Critical unknowns remain: pricing per call, the maximum number of candidate answers per request, and whether developers can fine-tune the model on proprietary data. These specifics will determine whether the API becomes a standard component in AI agent frameworks or remains a specialized tool alongside chatbots.
The company's existing Moderation API offers a parallel model, returning per-category scores rather than narrative text, though OpenAI predetermines those categories rather than allowing developer customization.
