Decisions API
OpenAI's API for questions with a bounded answer space, such as classification and request routing, with support for probability estimates, predefined options, and scores.
What changed
The Decisions API moves from public beta to general availability. By separating out small decisions such as classification and request routing, it lets developers evaluate performance within a clearly defined answer space.
Key readingAI-assisted analysis
Separating frequent small decisions into steps that can be evaluated independently helps developers pinpoint where delays and misjudgments occur. The actual efficiency gains depend on how those decisions connect to queues, review processes, and permission rules.
In focusOpenAI gives routine decisions a dedicated API
This breakdown starts with a story focused on this topic. It is not a complete assessment.
The facts
01From public beta to general availability
On October 22, OpenAI announced the general availability of the Decisions API, powered by GPT-6 Luna. The release adds prompt caching, native JSON input, and integration with GPT-Live, while raising the limit to 400 questions per request.
02Define the question and answer space first
Developers supply questions, context, and the relevant options or rating levels. The API returns an estimated probability that a condition holds, a predefined option, or a score calculated by weighting each ordered level by its probability, for use in tasks such as classification and request routing.
H/ACC insights
Make small decisions independently testable
H/ACC editorial interpretation: Separating repeated small decisions into a dedicated API lets developers compare response times and decision errors for a specific task, instead of assessing a system solely by how smoothly an entire conversation flows. Defining the answer space and keeping application records also makes it easier to categorize errors and replay the cases involved.
Take customer service routing. An application can record which queue a request enters, whether it is reassigned later, and when it is handed to a human. The model's output informs the decision, while the application sets acceptance criteria and permissions for subsequent actions outside the model. Connecting these steps creates the opportunity to turn faster decisions into smoother service.
What to watchAI-assisted analysis
- On the same set of real requests, how do median response time, 95th-percentile latency, and the misrouting rate change compared with the current routing method?
- When the question wording, input language, or distribution of categories changes, can the misrouting rate and the share of requests handed to a human stay within preset limits?
Limits of this assessment
- The performance figures in the public materials come from different tasks and comparison conditions. For this article, we have not called the API or independently replicated its speed, accuracy, or cost results.
- Probabilities and scores are model outputs. Whether they are sufficient to trigger an action automatically needs to be tested on real tasks, with the application defining rules for accepting results and handing cases off.