OpenAI launches general availability of Decisions API
OpenAI has made the Decisions API generally available with GPT-6 Luna, offering a dedicated endpoint for applications and agents that need to make bounded decisions from text, images and structured inputs. OpenAI says the service can make these decisions at up to 10 times the speed of GPT-6 Luna used through the Responses API.
The API is aimed at recurring workflow steps such as classifying an input, routing a request or choosing an agent’s next action. It supports up to 400 questions in a single request, twice the prior limit, and accepts native JSON as well as images in question instructions.
Three structured response types
Developers define a question, specify the permitted answers and provide the context needed to distinguish them. The Decisions API returns one of three response types: a predicate provides a probability from 0 to 1, a choice selects from supplied values and also returns probabilities and confidence, and a score produces a probability-weighted average across ordered rubric levels.
This model is designed for cases where an application needs a constrained result rather than an open-ended response. The general-availability release builds on OpenAI's Decisions API announcement by providing a dedicated endpoint for these narrowly defined decision tasks at production scale.
Since public beta, OpenAI has added prompt caching for reused context, inference residency in the US and EU, and further capacity in several Asian regions. Eligible customers can also use Zero Data Retention. Pricing is set at $0.10 per million input tokens, or $0.01 per million cached input tokens; OpenAI says there are no output-token charges.
Voice workflows and reported customer results
The API now integrates natively with GPT-Live for live voice experiences. In a GPT-Live session, transcript updates can trigger a choice such as an avatar reaction, while support applications can use the result to ask a clarifying question, route a request or escalate an issue.
OpenAI cited early customer evaluations across several workflows. Box reported approximately five-times-faster route selection in an evaluation of 30 certificate-of-insurance cases, with the same accuracy as the Responses API. Datadog reported 40-times-faster p95 performance than available frontier models with comparable accuracy for scoring the risk of agent-generated shell commands before execution.
Ramp said the service made AI cost attribution four times faster and 30% cheaper than GPT-6 Luna without sacrificing accuracy. Sierra reported under 280 ms p95, about three times faster and 90% lower cost per call than its previous approach in parts of live customer-conversation orchestration. SpeakBox reported reducing time to next action from 1.4 seconds to 0.13 seconds in testing with GPT-Live.
What teams should evaluate
OpenAI recommends starting with one recurring decision, defining its question, possible answers and necessary context, then testing against examples that reflect real use. For businesses, the practical implication is to validate accuracy, confidence thresholds, latency and reviewer escalation paths on a bounded workflow before relying on automated decisions in a live process.