OpenAI Adds an API for Choosing From Preset Answers, in Limited Preview
GPT-6 Luna powers the new endpoint, which accepts text or images as context. Broader access is planned in the coming days.
Loading page…
GPT-6 Luna powers the new endpoint, which accepts text or images as context. Broader access is planned in the coming days.
Listen to this story
OpenAI’s Decisions API gives applications a way to ask GPT-6 Luna to select from developer-defined answers, rather than generate an unrestricted response. The limited preview is restricted to selected API customers; OpenAI says broader availability is planned in the coming days but has not named a date. The fixed-choice design could help teams build classification, routing and agent-action workflows, but the announcement provides no latency, accuracy or pricing data, leaving those tradeoffs for preview users to evaluate.
Each request combines a question, a set of possible answers, and text or images for the model to evaluate.
OpenAI’s support example routes a request to a team; it does not generate a customer-facing reply.
The preset choices define what the model can return, but do not establish how reliably it selects the right option.
Selected developers can now test an OpenAI interface for directing incoming requests and choosing an AI agent’s next action. OpenAI Developers introduced the Decisions API on September 29, powered by GPT-6 Luna. The API—a way for software to request a model’s output—is in limited preview, with access restricted to selected API customers for testing.
Each request includes a question, possible answers and text or images as context. Luna returns a selection the application can use. That gives developers a defined answer set rather than an open-ended response: the question identifies the decision, while the context supplies the material to evaluate. Developers set the available choices before the model makes its selection.
OpenAI puts three jobs at the center of the product: classifying content, routing requests and choosing an AI agent’s next action. These cover different parts of an application: assigning a category to content, identifying a destination for an incoming request or selecting the next step for software acting on a task. The possible answers therefore depend on the job: categories for classification, destinations for routing, or actions for an agent.
OpenAI pitches the interface as offering real-time decision-making, but its announcement includes no response-time measurements, accuracy figures or pricing. Developers testing the preview will need to assess speed, reliability and cost for their own workflows. A preset answer set defines what the model can choose; it does not by itself show how well the model chooses.
OpenAI says a broad release is planned “in the coming days,” without specifying a date. For now, the endpoint remains available only to the selected API customers testing it in the limited preview.
Loading discussion...
Join the conversation
Share a situation where one approach would serve you better.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.