For years the pattern was always the same. You ask the model a question, it writes you a paragraph, and then you write code to dig the actual answer out of that paragraph. Is this a refund request? Which team should handle it? How urgent is it? The model would tell you in three sentences and your regex would pray it caught the right one.
OpenAI just shipped the thing that kills that whole dance. It is called the Decisions API, it is in public beta, and instead of generating text it answers typed questions. You send it a text or an image plus a list of questions, and it sends back answers with probabilities attached. A predicate question returns the probability that a condition is true. A choice question picks one of your fixed options. A score question rates against ordered levels you define. Any single question can come back as a refusal, and the rest still get answered.
Let me translate that into plain work. One request can now carry up to 400 independent questions, and the bill is input tokens only, ten cents per million, with nothing charged for output or caching. OpenAI says it runs about ten times faster than asking the same thing through the regular Responses API, which I would file under vendor claim until someone benchmarks it independently, but the direction is obvious. This is the model as a classifier with a confidence number, not the model as an essayist.
I keep coming back to how this fits the week I have been writing about. Two days ago I covered Anthropic's Duet, the always-on agent that works while you sleep, and I said the new front of the AI race is above your inbox, not inside a chatbot window. Decisions is the plumbing underneath that front. An agent that schedules, triages, and routes needs to make a thousand small judgments a day, and it does not need a paragraph for any of them. It needs a yes or no with a number on it, fast and cheap, so the expensive thinking can go where it matters.
My honest take, as someone who triages messages for a living
Here is my real-world version of this API. My Fiverr inbox gets messages from strangers every day, and a meaningful fraction of them are scammers running the same two plays: the fake order claim and the please-open-this-file pitch. I have learned to spot them in seconds. A Decisions API call could do that triage before I ever look: is this a scam pattern, yes or no, with what probability, and anything below a threshold goes to me for review. That is the correct way to use a tool like this, and the documentation says the same thing in drier language. The probabilities are estimates, not guarantees. One independent tester found them leaning low on choice questions and high on a yes/no question, so you set your thresholds on your own labeled data and route the uncertain cases to a human.
That is also the part worth saying out loud. A probability next to an answer does not make the answer right. It makes the uncertainty visible, which is better, but only if someone is watching the number. The teams that will get burned by this API are the ones that treat 0.92 as a fact instead of a bet. The teams that will win are the ones that build the fallback first: uncertain goes to review, certain goes through, and the log of both gets checked every morning like an employee's work.
I find this release more interesting than most model launches, because it is not a model launch. It is OpenAI admitting that a huge share of real AI work is not conversation at all. It is judgment, repeated at scale, with the prose stripped out. The chatbot era taught models to talk. The agent era is teaching them to decide. This week, the deciding got an API.
