Much of what an AI agent does all day is not writing but deciding: which team should get this ticket, is this message a refund request, should the next step be a question or an escalation. Each choice is trivial; thousands of them add up in latency and cost. OpenAI's Decisions API, now generally available, is a dedicated endpoint for exactly these bounded questions.
It runs on GPT-6 Luna. A developer defines a question and its allowed answers, supplies context as text, images or JSON, and gets back one of three answer types: a probability, a choice from a list, or a score. OpenAI says this is up to 10x faster than asking Luna the same thing through the general Responses API. Since the public beta it has added prompt caching for reused context, inference residency in the US and EU, native JSON input, images inside question instructions, and up to 400 questions in a single request, double the earlier limit. A new link with GPT-Live lets voice apps make decisions mid-conversation, such as whether to route a caller or ask a clarifying question.
Pricing is unusual. Input costs $0.10 per million tokens and cached input $0.01, and there is no charge for output at all, which suits an answer that is a single number or label. OpenAI cites internal uses, including categorising personal finance transactions and scoring press coverage by theme.
The idea is not unique to OpenAI. Liquid AI's d1 models, which we review, also answer many named questions in one pass with calibrated probabilities, and they are small enough to run on a device rather than through an API. For teams already on the OpenAI API, the Decisions endpoint is the easier path; for those who want to keep data local or avoid per-call fees, an open decision model is the comparison to make.
Some caution applies. The speed figure is OpenAI's own and is measured against its own slower endpoint, not against rivals. OpenAI publishes no accuracy or calibration data in the announcement, which matters when a probability drives what an agent does next. For a single fixed task with plenty of labelled examples, a small fine-tuned classifier may still be faster and cheaper. The endpoint is most useful where questions change often and training a new model each time is not worth it.