OpenAI Releases Decisions API in Public Beta, Powered by GPT-6 Luna

OpenAI released the Decisions API in public beta on October 6, 2026, introducing an endpoint that returns typed answers to developer-defined questions over text and image inputs. The OpenAI API changelog records the release in beta with gpt-6-luna, served through a dedicated POST /v1/decisions endpoint.

OpenAI says the new endpoint produces those answers about 10x faster than the Responses API. According to the Decisions API documentation, the API can be tested in the Playground, where developers can experiment with questions and inputs ahead of writing code.

The release follows the API’s introduction at DevDay 2026, where it was available in limited preview. OpenAI’s DevDay recap said the API applies the Luna model’s reasoning to a set of user-defined questions, each with a finite set of predefined answers, letting developers supply text or image context and get back results that can drive content classification, request routing, or an agent’s next action. OpenAI said at the time that a broad release was planned in the coming days.

The endpoint runs on GPT-6 Luna, a reasoning model OpenAI released alongside GPT-6 Sol on September 22, 2026. Both models accept text and image inputs and generate text through the Responses and Chat Completions APIs.

How the Decisions API Works

Each request has three parts. The model field selects the evaluating model; currently, only gpt-6-luna is supported. The input field carries shared evidence for the questions, either as a text string or as user messages containing text and images. The questions array specifies what the model should evaluate: each question’s type, its instructions, and any allowed choices or score levels. The response contains an answers array, and each question carries a unique name that the API echoes back with its answer.

A predicate question checks a condition, such as visible damage on a product or whether a passage is relevant, and returns a probability estimate from 0 to 1 that the condition is true. In the guide’s example, a product photo is inspected for a crack, tear, or dent, and an illustrative response returns a probability of 0.92; an application can flag photos for review once the value crosses a threshold the developer sets.

A choice question selects one value from a fixed set of supplied options, each with a description explaining when it applies. The answer contains the selected value, a probabilities array covering every option, and a separate confidence field. The guide’s routing example sends a customer complaint about a double charge to a billing option ahead of technical, shipping, and other alternatives. OpenAI’s guidance is to add a fallback value such as other for inputs the categories do not anticipate; the application can route that fallback outcome to a general review queue.

A score question rates an input against ordered levels, such as issue severity, arranged from lowest to highest. The returned score is the probability-weighted average of the level indices, which start at 0, so the outcome can land between two defined levels. In the worked example, probabilities of 0.1, 0.7, and 0.2 across three severity levels produce a score of 1.1 with a confidence of 0.55.

For outputs beyond these three answer types, the documentation points developers to Structured Outputs with the Responses API when they need an object that follows a custom JSON schema, and to function calling when a model must request a tool call with arguments.

Usage Guidance and Input Limits

Images must be sent as inline base64 data URLs; the endpoint does not support hosted HTTP or HTTPS image URLs or file_id inputs. To evaluate an image alongside instructions or other context, developers place input_text and input_image parts together in a single user message.

Independent questions can share a single request’s questions array, and each question can use a different type, so an application can check a product photo for damage and classify the product’s category in one call. When one decision depends on an earlier answer, the application must send separate requests, using the first result to gate the follow-up, such as confirming damage before asking for a repair category.

The documentation’s authoring guidance centers on observable criteria: keep distinct concerns in separate questions, give each choice a distinct meaning, and define score levels so that neighboring levels differ clearly. On interpretation, the guidance is to calibrate thresholds against labeled examples from the production application, weighing the relative costs of false positives and false negatives.

Pricing, Data Controls, and Voice

Pricing for the endpoint with gpt-6-luna is $0.10 per 1M input tokens. Only input tokens are billed: cache reads, cache writes, and output tokens carry no charge. Regional processing premiums and long-context input pricing multipliers still apply, while gpt-6-luna requests outside /v1/decisions follow the applicable model and processing-tier pricing.

The API supports Zero Data Retention and HIPAA use for eligible customers, with data residency and regional processing offered in the United States and in Europe (the European Economic Area and Switzerland). For voice-driven applications, client delegation with the Live API lets an agent choose actions from voice requests and report the results back to the user.

OpenAI says it expects the Decisions API to reach general availability in the coming weeks.