OpenAI launches Decisions API for fast classification and routing

OpenAI has introduced the Decisions API, a new tool designed to turn text and images into structured decisions for software applications. Released in public beta on October 6, 2026, this endpoint uses the gpt-6-luna model to evaluate inputs and return typed answers significantly faster than previous methods. The launch targets developers who need to classify content, route requests, or prioritize work without generating lengthy natural language responses.

What happened

The Decisions API is now available via a dedicated POST /v1/decisions endpoint. It is currently in public beta, with OpenAI expecting a general availability release in the coming weeks. The only model supported at launch is gpt-6-luna. To use the new features, developers must update their OpenAI SDKs to specific minimum versions: Python 3.26.0, JavaScript 7.30.0, Go 3.73.0, Ruby 0.101.0, and Java 4.78.0.

This API differs from standard chat completions by focusing on evaluation rather than generation. Instead of writing paragraphs, the model returns specific data types such as probabilities, categorical choices, or numeric scores. OpenAI states that this approach delivers answers about ten times faster than the Responses API. The service supports Zero Data Retention and HIPAA compliance for eligible customers, with data residency options in the United States and Europe.

How it works

A request to the Decisions API consists of three main parts: the model, the input, and the questions. The model field specifies gpt-6-luna. The input field contains the shared evidence, which can be a text string or user messages including images. The questions array defines what the model should evaluate, including instructions and allowed outputs. Each question must have a unique name, which the API echoes back in the response to help match answers to queries.

Developers can choose from three question types depending on the decision logic required. A predicate type checks if a condition is true, returning a probability between 0 and 1. This is useful for binary checks like detecting visible damage in a photo. A choice type selects one option from a fixed set, such as categorizing a support ticket by department. A score type rates input against ordered levels, such as issue severity. Score levels start at index 0, where 0 means cosmetic, 1 means a workaround exists, and 2 means fully blocked. The returned score is a probability-weighted average of these indices, allowing for nuanced results like 1.1.

For complex workflows, developers can include multiple independent questions in a single request. For example, a product photo can be checked for damage and classified by category simultaneously. However, if a decision depends on a previous answer, separate requests are recommended. OpenAI advises writing questions around observable criteria and setting thresholds based on the cost of false positives and negatives in your specific application.

Key details

• The Decisions API is approximately 10x faster than the Responses API for evaluation tasks.
• Input costs for gpt-6-luna on this endpoint are $0.10 per 1 million tokens.
• There are no charges for cache reads, cache writes, or output tokens on this endpoint.
• Regional processing premiums and long-context input pricing multipliers still apply.
• The API supports Zero Data Retention and HIPAA use for eligible customers in the US and Europe.
• Scores are calculated as probability-weighted averages of level indices, which start at 0.

Why it matters

For engineering teams building production AI features, latency and cost are critical constraints. Traditional large language models often generate verbose explanations even when a simple yes/no or category label is needed. The Decisions API strips away this overhead, providing structured data that integrates directly into application logic. This reduces the need for post-processing steps to parse natural language outputs, simplifying the codebase and improving reliability.

The ability to handle both text and images in a single evaluation step opens new possibilities for automated quality control and triage. Support systems can automatically route tickets based on urgency scores derived from customer messages. E-commerce platforms can flag damaged goods from user-uploaded photos without human review. By exposing confidence levels and probability distributions, the API allows developers to build more robust fallback mechanisms, ensuring that low-confidence decisions are routed to human agents.

What you can do

• Update your OpenAI SDK to the required version for your language stack before testing.
• Use the Playground to experiment with different question types and input formats.
• Define clear, observable criteria for each question to improve model accuracy.
• Set custom thresholds for routing decisions based on your tolerance for error.
• Combine multiple independent questions in one request to reduce API calls.
• Review data residency requirements if you operate in Europe or handle health data.