OpenAI
OpenAI's Decisions API bills only input tokens, at $0.10 per 1M
OpenAI's Decisions API beta returns typed answers from text and images on gpt-6-luna, bills input tokens only and is 10x faster than Responses.
OpenAI released the Decisions API, a beta endpoint that turns text and images into typed answers and charges only for input tokens. Developers who classify content, route requests or prioritize work pay nothing for output.
Beta Launched Oct. 6 With One Model
OpenAI added the API Oct. 6, 2026, according to the OpenAI API changelog, read Oct. 14. The entry says it returns answers “10x faster than the Responses API.”
The endpoint is POST /v1/decisions. The only supported model is gpt-6-luna. The docs guide calls the API a public beta and says OpenAI expects general availability “in the coming weeks.”
Input Tokens Are the Only Charge
OpenAI bills only input tokens, at $0.10 per 1M on gpt-6-luna, the docs guide said Oct. 14. There are no cache-read, cache-write or output-token charges. Regional processing premiums and long-context input multipliers apply.
OpenAI’s pricing page, read the same day, lists gpt-6-luna on the standard Responses tier at $0.10 per 1M input tokens and $0.50 per 1M output tokens for short context.
Requests Pair an Input With Questions
A request pairs an input with named questions and a model. Input is a text string or user messages with text and images. Each question has a type, instructions and any allowed choices or score levels. Each gets a unique name that the response echoes in an answers array.
Answers Come Back as Probabilities
Each question type returns a probability-based number. A predicate question gives the probability, from 0 to 1, that a condition is true, such as visible damage in a product photo.
A choice question picks one of the values the caller supplies, such as a department. It returns a probability distribution and a separate confidence field.
A score question rates the input against ordered levels, such as issue severity. It returns the probability-weighted average of the level indices. It can fall between levels. Probabilities of 0.1, 0.7 and 0.2 over levels 0, 1 and 2 give 1.1. A confidence field comes with it. Any answer can also be a refusal.
Thresholds Come From Labeled Examples
OpenAI tells developers to set thresholds for routing, filtering or review from labeled examples from their own application, weighing the cost of false positives against false negatives. A decision that depends on an earlier answer needs a separate request. Independent questions about the same input can share one.
For a custom JSON schema or written explanations, the docs point to Structured Outputs with the Responses API. For tool calls, they point to function calling.
HIPAA and Zero Data Retention Are Supported
The API supports Zero Data Retention and HIPAA use for eligible customers. Data residency and regional processing are available in the United States and Europe (the EEA and Switzerland).
The minimum SDK versions are Python 3.26.0, JavaScript 7.30.0, Go 3.73.0, Ruby 0.101.0 and Java 4.78.0. A Playground sits at platform.openai.com/decisions.
Our View
We think the billing rule matters more than the speed claim. On the Responses tier, gpt-6-luna output costs $0.50 per 1M tokens. A developer who sorts support tickets or screens product photos with Decisions pays only for what goes in.
The bill then follows the size of what the developer sends. The trade is fixed answer shapes. A developer who wants a written reason or a custom JSON shape must use Responses.
Developers must also set thresholds from their own labeled examples. A probability needs a cutoff before it can route anything. The savings last only if the questions fit fixed answers and the thresholds hold up on real traffic.