OpenAI Decisions API now available on AI Gateway
By Steven Van ·
One request can ask yes/no, choice and score questions about the same input for routing, triage, guardrails or rubric scoring.
Routing, triage, guardrails and rubric scoring can now use decision models through an OpenAI-compatible /v1/decisions endpoint on Vercel AI Gateway. These models return probabilities, choices and scores instead of generated text.
Ask several questions about one input
A single request can ask predicate (yes/no), choice and score questions against the same input. For example, the release note shows a refund request being assigned to either a billing or technical team. OpenAI SDK answers arrive as an array in question order, with each answer carrying its question's name. In the AI SDK, questions are keyed by name, each key becomes a key in answers, and the yes/no type is called boolean.
Choose a decision model
The openai/gpt-6-luna-decisions model ID is separate from openai/gpt-6-luna, which remains a language model for text generation. Requests to the decisions ID use the Decisions API and are billed at its rates. Other supported decision models include Jev, Liquid d1 and Laya, alongside GPT-6 Luna Decisions in the model directory linked from the announcement.
Call through an SDK, HTTP or CLI
Developers can point the OpenAI SDK at AI Gateway and call decisions.create, or use the AI SDK, the existing /v1/evaluate HTTP API or the AI CLI. The OpenAI SDK requires version 7.30.0 or later for JavaScript, or 3.26.0 or later for Python. The AI SDK's experimental_decide requires ai 7.0.128 or later.