OpenAI Decisions API Beta Offers 10x Faster Typed Answers

OpenAI has launched the Decisions API in public beta, a new tool designed to translate text and image inputs into structured, typed answers that can be directly used by code for branching logic. The company states that this new API operates approximately 10 times faster than its existing Responses API, addressing a common developer pattern of prompting a large language model and then parsing its textual output into specific labels or decisions. The Decisions API aims to streamline this process by providing pre-typed, actionable outputs.

The Decisions API supports a single model currently, identified as gpt-6-luna, which features a substantial 1,050,000-token context window. Access to the API is exclusively through OpenAI's hosted endpoint via a POST request to /v1/decisions, with no options for open weights or self-hosting. Pricing for the Decisions API is set at $0.10 per 1 million input tokens, with no additional charges for output, cache reads, or cache writes, positioning it as a cost-effective solution for rapid, typed decision-making. OpenAI anticipates general availability for the API in the coming weeks.

Each request to the Decisions API includes three primary fields: model, input, and questions. The input can comprise text, images, or a combination of both, serving as the evidence for evaluation. The questions field is a list, where each question is defined by a unique name, a specified type, and accompanying instructions. The API's response returns an array of answers, keyed by the names provided in the request. This structured output format is intended to simplify integration into downstream applications.

The Decisions API supports three distinct question types: predicate, choice, and score. The predicate type evaluates a condition and returns a probability between 0 and 1, such as determining if a product photo displays a defect. The choice type selects one value from a list of user-supplied options, also providing per-option probabilities and an overall confidence score. The score type rates the input against ordered levels, returning a probability-weighted average of level indices. For instance, OpenAI illustrates this with a severity example where probabilities of 0.1, 0.7, and 0.2 for different levels result in a score of 1.1, indicating a specific severity category. OpenAI distinguishes the Decisions API from its Structured Outputs feature, recommending Decisions for probabilistic outputs, choices, or scores, while Structured Outputs are for prose generation.