OpenAI's Decisions API is generally available at $0.10 per 1M input tokens

OpenAI said in its API changelog on 22 October 2026 that the Decisions API is now generally available with gpt-6-luna, the only model it currently supports. The endpoint, POST /v1/decisions, takes text, images or JSON and returns typed answers instead of generated text. OpenAI's guide says it is about 10x faster than the Responses API.

What it returns

A request carries shared input plus a list of questions, and each question has one of three types. A predicate returns the probability that a condition is true. A choice picks one value from a list you supply, such as a department. A score rates the input against ordered levels, such as issue severity, as a probability-weighted average. Choice and score answers also carry a confidence field, and OpenAI tells teams to set thresholds from their own labelled examples, based on what a false positive and a false negative cost them.

General availability adds prompt caching, images inside individual question instructions, structured data passed as input_json, and a limit of 400 questions per request. OpenAI also lists a usage-reporting endpoint, a native link to its GPT-Live voice model, and Private Link support.

Pricing and data controls

According to OpenAI's guide, uncached input costs $0.10 per 1M tokens and output tokens are not charged. Cache hits are discounted and cache writes are charged, at the standard gpt-6-luna rates. Regional processing premiums and long-context multipliers apply. The implicit cache covers the shared input only if that prefix is at least 1,024 tokens; question text sits outside it. OpenAI says the API supports Zero Data Retention and HIPAA use for eligible customers, and regional processing in the United States and Europe (EEA and Switzerland).

What customers and rivals say

TechCrunch reports that OpenAI's Nikunj Handa described the product as a reconfigured Luna rather than a purpose-built classifier, and said latency dropped from 1 to 3 seconds to 100 to 180 milliseconds. TechCrunch also reports that Speak's CTO, Andrew Hsu, said the time between actions fell from 1.4 seconds to 0.13 seconds, and that Box, Datadog and Ramp tried the API in early trials. Those are OpenAI's and its customers' accounts, not independent measurements.

The catch is that fast is not the same as best. TechCrunch says two public decision-model indices, the AIM System One Index and the Artificial Analysis Decision Index, favour Liquid's D1 Ultra and Base Lab's Elephant-1, and place OpenAI's model eighth and fourteenth. It adds that TypeSafe's Jev models are still the most queried on OpenRouter, and quotes TypeSafe's CEO urging people to trial models on their own work. OpenAI's own page does not publish accuracy figures.

What to do

• List the classification, routing and triage steps in your agents that currently call a full language model, and note their latency and monthly input tokens.
• Build a labelled sample of 200 to 500 real cases per step and compare decision models on accuracy and calibration, not only price and speed.
• Check your data terms first: ZDR, HIPAA and regional processing are available only to eligible customers under the agreements OpenAI references.
• Keep a review path for low-confidence answers; OpenAI leaves the threshold to you.

Not stated in the sources: gpt-6-luna's accuracy on any named benchmark from OpenAI, the cache-read price per token, and whether other models will be added.