OpenAI Launches Decisions API, Says It Makes App Decisions Up to 10x Faster
The developer service runs GPT-6 Luna on decisions whose answer space is defined in advance. OpenAI says it returns decisions up to 10 times faster than using the same model through the existing Responses API.
OpenAI formally launched the Decisions API on October 22 local time. The service handles the work of classifying incoming requests or picking the next action in an app, letting developers wire GPT-6 Luna directly into their own products. It is built to process only decisions whose range of possible answers is set in advance. OpenAI says it reaches a decision up to 10 times faster than running the same model through the existing Responses API.
What it handles inside an app
It is meant for deciding where to route a request and which action to take next. Inputs can include text, images, and data structured in a predefined format. A single request can carry up to 400 questions at once, and images can be embedded within those questions.
Results come back in one of three forms, depending on the use case. The first returns the probability that a given condition is true as a number between 0 and 1. The second asks the model to pick one option from a predefined list, returning probability and confidence alongside it. The third returns a score that factors in the probability assigned to each level in an ordered set of evaluation levels.
How it works in live conversations
Since the public preview, direct integration with GPT-Live has been added, allowing decisions to be made on the spot in the middle of a real-time conversation. OpenAI offered the example of selecting an avatar’s facial expression to match what is being said. The update also adds the ability to store and reuse repeated inputs, along with processing in the United States and the EU.
How much faster in practice
Customer accounts center on shorter waits before an app can move on to its next action. Speak said the time to reach the next action dropped from 1.4 seconds to 0.13 seconds. Box, evaluating 30 certificates of insurance, said the time needed to choose a routing path fell to roughly one-fifth of what it was.
Some cases point to improvements in both cost and processing speed. Ramp said that on the job of allocating AI costs to actual work, the Decisions API ran 4 times faster and 30% cheaper than GPT-6 Luna. Sierra said the cost per call dropped by 90%.
What it costs
Base input pricing is $0.10 per million tokens. Reusing stored inputs costs $0.01 per million tokens, and there is no charge on the output side. The write cost for storing inputs for either 30 minutes or 12 hours is listed separately in the pricing documentation. Anyone planning heavy use of repeated inputs will need to factor that cost in as well.
Developers can try the features out in advance through the Decisions Playground and the developer guide. Zero data retention is available only to qualifying customers. Teams applying this to Korean-market services would do well to weigh decision accuracy on Korean-language inputs, real-world latency, and total cost including input storage fees.