OpenAI announced general availability of the Decisions API on October 22, offering applications and AI agents a service for classifying information or quickly choosing between defined actions. The product uses GPT-6 Luna and accepts text, images or JSON data. Its response is bounded by the application's question: a probability, an option from a list, or a score.
According to OpenAI, this approach can produce decisions up to ten times faster than using GPT-6 Luna through the Responses API. The comparison comes from the company and concerns these two ways of accessing the model. The announcement builds on the beta with caching for reused context, native GPT-Live integration, regional processing and more questions in a single request.
Three answer types for application workflows
The Decisions API is built around questions whose possible results are defined in advance. To check a condition, the type called predicate returns an estimate between zero and one of the probability that the condition is true. A documented example examines a product photograph for visible signs of damage, using the criteria included in the question.
The choice type selects one of the options supplied by the developer. In a support system, a request can be assigned to a department, while an agent can receive a next step from options such as continuing, retrying or seeking clarification. The response includes the selected option, probabilities for the alternatives and the model's confidence indicator.
The score type evaluates information against ordered levels in a rubric. Its result is the average of those levels' indices, weighted by the probabilities assigned by the model. The value can fall between two levels. This allows an issue's severity to be expressed on a defined scale, while selecting a single category remains the role of the choice type.
The application defines the questions, permitted options and information needed for evaluation. The application also executes the operations associated with an answer through its own logic. Tool calls and their outputs can be included as evidence in a conversation being evaluated. The Decisions API returns the classified result, which the software workflow uses for routing, filtering or prioritization.
JSON, images and up to 400 questions
General availability adds support for JSON objects and arrays passed directly as evidence to analyze. Structured data can be combined with text and photographs in a message. Images can be supplied through public URLs or as data included directly in the request. Each question's instructions can also contain their own images, alongside the evidence shared across the evaluation.
The limit is now 400 questions per request, up from 200 during the beta. Independent questions can use the same context, including for different evaluations of a single photograph. Where a question depends on an earlier answer, the documented mechanism uses successive requests. This preserves both the shared evidence and the sequence of decisions within the application.
Reused context benefits from prompt caching, allowing an identical prefix of input evidence to be used again in subsequent requests. That prefix must contain at least 1,024 tokens. Question content sits outside this reusable portion, and images specific to a question follow the same boundary. The division determines which material can benefit from the cache on later calls.
Integration with voice applications and document processing
Native GPT-Live integration allows evaluations to be triggered during a voice conversation as its transcript updates. OpenAI describes a support agent choosing between asking another question, routing the request and escalating the case. In the company's digital character example, the application receives the selected reaction and changes the character's expression accordingly.
Voice conversation processing remains in GPT-Live, while the Decisions API evaluates the evidence supplied to select a result. The mechanism connects conversation events to actions implemented in the application. This integration is among the capabilities added since beta, alongside expanded processing capacity in several Asian regions, which OpenAI mentions in its announcement.
Box is among the companies describing use of the product to classify documents and send them to the appropriate next stage. In Box's reported evaluation of 30 certificate-of-insurance cases, route selection was approximately five times faster than through the Responses API, with the same accuracy. The company says lower-confidence cases are sent to human reviewers.
OpenAI also describes use within its own teams. Personal Finance uses the product to categorize transaction descriptions. Its communications team classifies coverage by subject and evaluates its relationship to the narratives being tracked, helping identify material that needs closer examination.
Input-based pricing and regional options
The announced base price is $0.10 per million input tokens. Reading cached input costs $0.01 per million tokens, while cache writes have separate charges based on retention duration. Output tokens carry no charge. Regional processing and long-context requests have the pricing adjustments specified in the service's conditions. These are distinct elements of the endpoint's billing.
Regional processing is supported in the United States and Europe, with the latter covering the European Economic Area and Switzerland. Zero Data Retention is available to eligible customers. OpenAI distinguishes these conditions from simply using the Decisions endpoint, for which GPT-6 Luna is currently the supported model. Eligibility requirements apply to the relevant data controls.
Businesses can separately monitor token consumption and request counts, filtering or grouping them by project, API key or model. A request containing several questions is counted once in this metric. General availability therefore brings classification, image evaluation and next-step selection together in a service dedicated to structured decisions within applications.
