OpenAI made its Decisions API generally available on October 22. The endpoint runs on GPT-6 Luna. It returns bounded answers in the form of probabilities, choices and scores. Product lead Nikunj Handa said the team reconfigured ChatGPT Luna to speed up inference. The change also lowered cost. Handa told reporters the API reduced processing times from one to three seconds. The new range is 100, 120 or 180 milliseconds. The routing decision arrives before the user notices the wait.
The construction ERP routes payroll webhooks through a gate when a job closes. Atelier sends inference calls to seven providers and logs the decision to D1. The estimating tool checks dollar amounts and queues a review when the threshold breaks. Each gate asks a question and waits for an answer. The wait used to be 1.6 seconds. OpenAI says the Decisions API hits 150 milliseconds.
The API returns a selection from a list
The Decisions API is designed for high-volume, low-latency decision-making. It returns typed answers that applications use to classify content, route requests and prioritize work. One request can carry up to 400 questions. Output costs nothing. Input costs ten cents per million tokens.
The flight planner checks NOTAMs before every leg. The permit app classifies a photo as deck or porch. The dating-profile teardown scores a bio on a rubric. Each call sends context and gets back a label. The label tells the system what happens next. Early customer Speak replaced a language-model call with the Decisions API. The time between actions dropped from 1.4 seconds to 0.13 seconds. The API picked faster.
The latency claim is a target
Handa framed the forward target as a 60-millisecond call. That is fast enough for an API call for every single frame in a video. OpenAI product lead Tibo Sottiaux described the current performance as less than a few hundreds of milliseconds end to end. That is a tuning target. The API costs 2.4 times what TypeSafe charges for its rival Jev model. The price is higher. The promise is multimodal.
The construction ERP closes the books at month-end. The payroll gate logs every decision and timestamp. If the decision arrives in 150 milliseconds the gate opens and the webhook fires. If the decision takes three seconds the user sees a spinner. The user wonders if the system hung. General availability landed on October 22, a little over three weeks after the first preview. Three weeks ago the call took 1.6 seconds and now it might take 150 milliseconds.
The decision arrives before the user finds the refresh button.
Every gate in production trusts that the model will answer. The timestamp says when the answer arrived. Handa said OpenAI might introduce specialized training to improve these capabilities in the future. The current product uses Luna without that training. The model that writes the paragraph also picks the route. It just does it faster now.