OpenAI’s Decisions API picks an answer in 150 milliseconds

OpenAI on September 29, the day of its DevDay conference, opened a limited preview of the Decisions API. It takes a question, a fixed set of answers and some context, then returns a confidence score for each answer instead of a chat reply, and the company’s post on Twitter says it accepts images and is tuned to decide in under a few hundred milliseconds end to end. If your team retrains a small classifier every time its label set changes, or squeezes a confidence score out of a chat model’s token probabilities, the Decisions API takes the labels with the request and scores them directly, though for now only in preview.

OpenAI puts the response time at 150 milliseconds, against 1.6 seconds for GPT-6 Luna, the cheapest model in its lineup and the one the Decisions API is built on, The New Stack reported. The developer supplies the questions, the candidate answers and the context. What comes back is the predefined answers with their scores, which suits classifying content, routing requests or picking an agent’s next action from a short list.

Most teams handle that job today with a regular chat model and a carefully worded prompt, The New Stack notes. A chat model’s confidence is often a rough guess, and it burns a fair number of tokens getting there. The usual alternative, a small trained classifier, is fast and cheap, but it needs labeled data and a fresh training run whenever the labels change. A decision model sits between the two: new labels go in the prompt, and the score it returns is one a developer can act on.

OpenAI already sells a narrower version of the idea. Its Moderation API has long returned per-category scores rather than prose, but there OpenAI sets the categories, not the developer.

The model underneath has public rates: GPT-6 Luna, released on September 22, costs $0.10 per million input tokens and $0.50 per million output tokens. Those are Luna’s token rates, though, and they don’t price a Decisions API call. That price, how many candidate answers one request can hold and whether the API can be tuned on your own data are all still open. The New Stack says those details will decide whether the Decisions API becomes a standard building block in agent frameworks or stays a niche tool next to the chat models.

The publication reads the launch as a likely reply to TypeSafe’s Jev, whose sudden rise made decision models popular, and thinks OpenAI probably rushed the announcement out ahead of DevDay. An OpenAI spokesperson said the company plans to share more “at broad rollout.” OpenAI plans that broad release for the coming days.