OpenAI has made its Decisions API generally available, introducing a dedicated interface for fast, structured decision-making in AI applications. The release targets a common challenge in AI development: applications and autonomous agents often need to make numerous small decisions throughout a workflow. Classifying incoming information, routing requests, evaluating risk, and selecting the next action may each require little reasoning, but their cumulative latency and inference costs can become significant at scale.
The Decisions API is designed to handle these tasks efficiently, allowing developers to integrate structured AI decisions directly into application logic.
Unlike general-purpose interfaces designed to handle open-ended conversations and complex tasks, the Decisions API focuses on questions with clearly defined answer formats. Developers specify a question, the available answers or evaluation criteria, and the context needed to make a decision. The API then returns a structured result that an application can use immediately.
The API supports three decision types:
- Predicate: Estimates the probability that a condition is true. For example, a system can assess whether a product image shows visible damage.
- Choice: Selects one option from a predefined set. An AI agent could decide whether to continue a task, retry an operation, or request clarification.
- Score: Evaluates an input against an ordered rubric, such as assigning a severity level to a reported issue.
The API accepts text, images, and native JSON inputs, making it suitable for both language-based workflows and multimodal applications. Developers can also submit up to 400 questions in a single request, enabling multiple related evaluations to be processed together.
The general availability release introduces several improvements intended to make the API more practical for production environments:
- Lower latency: OpenAI reports decision-making speeds of up to 10 times faster than GPT-6 Luna through the Responses API.
- Prompt caching: Reusing shared instructions and context can reduce the cost of repeated requests.
- Native GPT-Live integration: Applications can trigger decisions during real-time voice interactions, enabling more responsive conversational experiences.
- Expanded multimodal support: Developers can pass structured JSON directly and include images in individual question instructions.
- Regional processing: Inference residency is available in the United States and the European Union, with Zero Data Retention available to eligible customers.
These capabilities make the API particularly relevant to systems where decisions must be made frequently and quickly, rather than only at the end of a longer AI workflow.
The API is designed for AI agents, customer support systems, document processing, and security monitoring. For example, voice assistants can respond more quickly, document platforms can automate classification, and security systems can evaluate commands before execution.
OpenAI reports significant improvements in customer deployments. Speak reduced the time to the next action in its language-learning experience from 1.4 seconds to 0.13 seconds in testing. Box reported approximately five times faster document-routing decisions, while Datadog achieved 40 times faster 95th-percentile latency in its evaluation of command-risk assessment.
These results reflect individual customer evaluations and may vary depending on the workload and implementation.
The Decisions API is generally available through OpenAI’s dedicated API endpoint, using GPT-6 Luna.
For developers building AI agents, conversational assistants, classification systems, and automated business processes, faster structured decisions can improve responsiveness while helping control inference costs. The approach is particularly useful when an application needs predictable output formats that can be connected directly to programmatic logic.
