The Decisions API is an OpenAI endpoint that does not write text: you give it a text, or text and images, plus typed questions, and it returns decisions with probabilities. It can say whether a condition holds, pick between options you provide or score against levels.
A question from the app. Pick an answer.
Jev is an AI model that does not write: it picks an option or gives a score, with probabilities.
You want a ticket's team, urgency and frustration. How, with fewer calls?
They look at the same text and do not depend on each other: they go together. If one needed another's answer to be built, that would take two calls.
It works, but you send the same ticket three times. Together they give the same answers: each question is answered on its own.
Within one call, answers are not passed between questions. If you really need the team to build another question, that is two calls.
It is an API: you send a request to POST /v1/decisions with model, input and questions. You can give each question a name to identify it in the response. There are three types, and a support email works as an example.
Does it ask for a refund? That is a predicate: it returns a probability from 0 to 1 that it holds.
Which team gets it? That is a choice: it picks one of the values you give it and returns choice, the probabilities of each option and a confidence.
How urgent is it? That is a score: ordered levels starting at 0, and it returns score, the average of the levels weighted by their probabilities (it can fall between two), plus probabilities per level and confidence.
All three go in the same request because they do not depend on each other. If one needs another's answer, those are separate requests. Images go inside the request, as a base64 data URL: an http link or a file_id will not do. How to read the numbers is in calibrated probability. A question can also return refusal, with no decision or probabilities.
According to OpenAI, the Decisions API is for getting probabilities, choosing between fixed options or scoring with a rubric. If you need your own JSON schema, say to extract fields or write an explanation, OpenAI points to Structured Outputs in the Responses API. And if the model has to call a tool, function calling.
The core difference: an LLM writes, token by token. Here there is no text to read, only values your code uses to decide what to do.
As of October 8, 2026, it is in public beta and only supports gpt-6-luna. It costs USD 0.10 per million input tokens. Output tokens are not charged, nor are cache reads or writes, though there are regional and long context surcharges. To run the numbers, see AI cost.
OpenAI says it is about 10 times faster than the Responses API. The guide does not explain how that comparison was measured. For eligible customers it offers ZDR (zero data retention) and HIPAA, with data residency in the US and Europe.
Its output resembles that of Jev, by TypeSafe AI: decisions with probabilities. TypeSafe uses System One for models like Jev, which decide instead of writing. The names differ: the yes or no question is called Noul in Jev and predicate here. If you switch from one to the other, test your thresholds again on your data.