The News, Explained
On October 22, 2026, OpenAI announced general availability of the Decisions API. An API lets an app request another service’s capabilities. Decisions currently uses GPT-6 Luna to evaluate text, images and JSON, a format that pairs named fields with values, and return answers within a defined range. Source Source
Three answer types serve different purposes. A predicate estimates a condition’s probability from 0 to 1. A choice selects a supplied option. A score averages ordered level indices using their probabilities, so it can fall between levels. Routing a request to a department fits a choice; rating severity against ordered criteria fits a score. The guide directs apps needing a written explanation or custom structured object to Structured Outputs through the Responses API. Source
Up to 400 independent questions can evaluate shared evidence in one request. Questions depending on earlier answers require separate requests. Prompt caching can reuse an identical shared-input prefix of at least 1,024 tokens, the units used to account for input. Individual question content sits outside that reusable prefix. Source
Native GPT-Live integration lets the voice service call Decisions directly or react to incoming transcript updates. The app executes the selected action. In the official car-assistant example, it raises the temperature, reads the actual setting and reports that confirmed state. The app connects a selected action with evidence that it happened. Source
Standard short-context input costs US$0.10 per million tokens; cache reads cost $0.01. Decisions has no output-token charge. Cache writes are charged, and regional-processing premiums and long-context multipliers apply. Estimating a request’s cost therefore involves the size and reuse of shared evidence alongside the question count. Source Source
Default abuse-monitoring logs are retained for up to 30 days. Eligible customers’ Zero Data Retention also has limitations: prompt caching may retain encrypted processing state on graphics processing units (GPUs) for up to 24 hours, and image inputs have a child sexual abuse material detection retention exception. Regional inference processing is supported in the US and Europe’s EEA and Switzerland. API availability and the location of inference are separate conditions. Source
OYOPICK’s Take
We see the value depending on how an app defines its questions. Routing inquiries to billing or shipping calls for clear category boundaries and a way to handle inputs fitting neither. Designing a review route for an intermediate score could make the connection between a brief decision and the resulting workflow clearer.
Whether faster classification improves a service would need to be judged against misrouted real inquiries and their outcomes. Voice apps also make changes easier to understand when they connect a selected action with its confirmed result. We would welcome this focused decision-making capability leading to clearer feedback about what actually happened.