The Pulse
OpenAI’s Decisions API Reaches General Availability, Raises Limit to 400 Questions
OpenAI’s Decisions API is now generally available and supports up to 400 questions per request, with GPT-6 Luna as its only supported model. The endpoint returns typed answers for predicates, choices and scores, and charges for input tokens
AI.info Team · Written with AI · fact-checked ·
OpenAI’s Decisions API is now generally available, offering typed answers for classification and scoring tasks. The endpoint supports up to 400 questions per request and currently runs only on GPT-6 Luna. OpenAI describes it as 10 times faster than the Responses API, though its documentation does not provide a benchmark method alongside that claim.
A separate endpoint for bounded answers
Developers send shared evidence and a set of questions to POST /v1/decisions. Each question can check whether a condition is true, select from supplied options or score against ordered levels. The response returns an estimated probability for a predicate, a selected value and probability distribution for choices, or a probability-weighted average of level indices for scores. A score can fall between the supplied levels.
“The Decisions API evaluates text, images, and JSON and returns typed answers about 10x faster than the Responses API.”
OpenAI’s API documentation
The format can be used for tasks such as routing support tickets, checking an image for visible damage or assigning a severity level. Teams can ask multiple independent questions about the same evidence in one request. Decisions that depend on an earlier answer require a separate request.
400 questions, images and structured input
OpenAI’s documentation says the limit is now 400 questions per request, up from 200 during the public beta. It also lists prompt caching, direct JSON input, images attached to individual questions, publicly hosted image URLs, conversation history that includes tool calls and outputs, and organization usage reporting among the capabilities added since the beta. Developers can also connect Decisions to GPT-Live, so an application can request a decision while a user speaks or as a transcript arrives.
User messages can include text, images and JSON; the documentation does not support files or audio as direct input. Each request uses GPT-6 Luna, with no model choice currently available.
Typed output is not a custom schema
The separation matters for teams already using the Responses API. Decisions is designed for predicates, fixed options and ordered scores; it does not replace a general structured-output workflow. OpenAI directs developers who need an object matching their own JSON schema, such as extracted fields or a written explanation, to Structured Outputs with Responses instead.
That boundary gives developers a choice between a limited set of answer types and a more flexible generation interface. A choice question requires developers to supply the possible values. The model can return a refusal for an individual question, and OpenAI advises teams to set routing thresholds using labeled examples and the costs of false positives and false negatives.
Input-token billing
For GPT-6 Luna, uncached input costs $0.10 per million tokens. Cache hits receive discounted input pricing, but cache writes are charged at the model’s standard prompt-caching rates. Output tokens are not charged. Regional processing premiums and long-context input multipliers can also apply.
The price and speed claims make the endpoint an option for frequent, bounded judgments, not a drop-in answer for every model call. Teams still need to evaluate whether its probability estimates suit their application, set thresholds for when a result triggers action or review, and account for the single-model constraint. The documented limits include 400 questions per request, GPT-6 Luna only, and three answer types: predicates, choices and scores.