An application deciding where to route a request may need one of three labels rather than a paragraph of explanation. OpenAI's Decisions API, generally available October 22, is designed for that kind of bounded question.
The launch announcement says developers define the question and allowed answers, then supply text, images or native JSON. The API supports three answer types and permits up to 400 questions per request.
Match the answer to the application
| Type | Result | Example to evaluate |
|---|---|---|
| Predicate | An estimated probability from 0 to 1 | Whether an input meets a defined condition |
| Choice | One of the supplied options | Which queue should receive a request |
| Score | A value across ordered rubric levels | How strongly an item matches a written standard |
A label can select a queue. A score needs an interpretation. A probability needs a threshold and a plan for cases that do not clearly meet it.
OpenAI's developer guide supplies the implementation details for those response types. Its launch rates are $0.10 per million input tokens and $0.01 per million cached input tokens, with no output-token charge. At the uncached rate, ten million input tokens would cost $1 before other services or applicable charges. This calculation describes token billing, not the cost of a completed support case.
The maximum of 400 questions in a request is a batching limit. A team still needs to decide which context belongs with each question and how it will check an answer. Sending more questions together cannot resolve an unclear routing policy.
Evaluation needs previously reviewed examples, including ambiguous cases and requests that should go to a person. Comparing the output with the expected handling can expose both routine routing mistakes and errors with more serious consequences.
For example, a routing trial could use three allowed values: billing, technical support and human review. A message mentioning both a payment and a broken feature could fit more than one queue. The expected label depends on the routing policy, which needs to be clear before the result can be judged.
OpenAI also announces a native integration with its live voice system and inference residency in the U.S. and EU. Those features have their own account and configuration requirements; they do not change the need to review the application's actual input and output path.
The application also needs a fallback for wrong answers, even when the API returns a valid label.
