Routing a support ticket or picking an AI agent’s next step rarely calls for a long answer, and OpenAI now sells a service that returns only short ones. The company made its Decisions API generally available Thursday, aimed at the labeling, routing and scoring calls that apps and agents make over and over inside a single workflow.
Decisions runs on GPT-6 Luna and handles questions with bounded answers, drawing on text, images and structured input. OpenAI said it is up to 10x faster than putting the same model to work through the Responses API. Requests go to a dedicated endpoint, POST /v1/decisions, and gpt-6-luna is the only model it supports for now, the developer guide says. Luna was also one of three OpenAI models that entered private preview in Google Cloud’s Model Garden on Wednesday.
Three answer types, and input is the only thing billed
A developer writes the question, lists the allowed answers and passes in the context. Predicate questions come back as a probability between 0 and 1 that a condition holds, such as whether a product photo shows visible damage. Choice questions return one of the developer’s own values along with probabilities and a confidence figure. Score questions grade input against an ordered rubric of 2 to 10 levels and return a probability-weighted average of the level indices, so a result can land between two levels.
The price is $0.10 per million input tokens and $0.01 per million for cached input, with no charge for output tokens. The guide adds some fine print: cache writes are billed, and regional processing premiums and long-context multipliers still apply. Zero Data Retention is open to eligible customers, as is HIPAA use.
What got added after the beta
Since the public beta, OpenAI has added prompt caching for reused context, inference residency in the US and EU, native JSON input and images inside individual question instructions. A request can now carry up to 400 questions, double the 200 allowed during the beta. Caching applies only once the shared input prefix reaches 1,024 tokens.
There’s also a native hookup to GPT-Live, OpenAI’s voice model. A Live session can fire a decision as transcript text streams in; OpenAI’s sample uses that to pick an avatar’s facial expression while the user is still talking. Andrew Hsu, CTO of language-learning app Speak, said the GPT-Live integration “was able to reduce the time to next action by over 90%, from 1.4s to 0.13s.”
The early speed claims come from customers
Datadog uses Decisions to rate the risk of agent-generated rshell commands before running them, so its agents can fix issues on their own. “Compared with available frontier models, it delivered 40x faster p95 with comparable accuracy—making it possible to place Decisions in the hot path of a remediation action,” said AJ Stuyvenberg, a staff engineer at the company.
Sierra picks next actions in live customer conversations with it. Engineer Ajeet Grewal put the latency at under 280 ms p95, which he said was about 3x faster, with 90% lower cost per call, than the company’s previous approach. At Ramp, research engineer René Sultan said attributing AI costs to the underlying work became 4x faster and 30% cheaper compared with GPT-6 Luna. Box measured roughly 5x faster route selection on 30 certificate-of-insurance cases. All of these figures are the customers’ own, published in OpenAI’s post.
The guide is specific about where the service stops. Decisions returns typed answers, and the calling application carries out whatever action an answer selects. For generated objects that follow a custom JSON schema, such as extracted fields or a written explanation, OpenAI points developers to Structured Outputs in the Responses API.