Microsoft's Decision-1 turns the check before an agent acts into a scored Foundry call

The interface is a closed question. An application sends a situation and the list of allowed answers, and Decision-1 sends back a probability for each one ^[2]. Input is text only, up to 32,768 tokens, and the response is JSON ^[4]. RuntimeWire's headline calls it a 9B model ^[3]. The supported formats are yes-or-no, multiple choice, ratings, rubric grading and relevance judgments, plus an explicit "cannot tell" option ^[5]. It does not write explanations or take open-ended questions, and it does not accept images, audio or video ^[6]. Srivastava's October 9 announcement lists its jobs as routing, classification, prioritization, verification and workflow control ^[17].

The "cannot tell" option is the best decision in the listing. A scorer that must split all its probability between "proceed" and "reject" has nowhere to put an input it does not understand. With an abstain choice available, escalation logic gets simple. RuntimeWire describes the pattern: an agent asks whether a proposed tool call meets a rubric, then proceeds, retries or escalates based on the score ^[7]. A team can send any case where "cannot tell" crosses a set level straight to a person.

Serial latency is the constraint behind the product. Agent checks are usually dependent, so step twelve waits on the verdict for step eleven. Microsoft's engineers give the figure in the announcement: 100 milliseconds added to each of 20 dependent decisions adds two seconds to a workflow ^[8]. Microsoft's median-latency figures put Decision-1 ahead of GPT-6 Sol by a factor of 35 and ahead of Quyet-1.0-Large by a factor of 4.5 ^[21]. A small scorer beating a general text generator on a task whose output is a few numbers is the expected direction of that result.

Input costs $0.042 per million tokens, and output is free ^[9]. Free output is an easy promise for a model that replies with a short list of probabilities. A check that fills the whole 32,768-token window costs about $0.0014 ^[19]. A thousand of those cost about $1.38 ^[20].

Across 36 benchmarks totaling nearly 150,000 questions, none of which the model saw in training, Microsoft says Decision-1 was the most accurate of the models it compared ^[22]. RuntimeWire notes that the announcement leaves out the full datasets and the latency test conditions, so outsiders cannot reproduce the comparison from it ^[10]. For the accuracy figure to carry over, a team's decisions have to resemble those benchmarks in their label sets, their prompt lengths and how often the right answer is actually ambiguous. For the latency multiples to carry over, the test's prompt sizes and concurrency have to look like production traffic.

Accuracy and calibration are separate properties. Accuracy counts how often the top choice is right. An escalation threshold needs a score of 0.8 on "proceed" to be right about 80 percent of the time on the team's own traffic. RuntimeWire frames the open question the same way: whether the score is dependable on customer-specific decisions, and calibrated well enough to set escalation thresholds safely ^[11]. A labeled sample of a team's own decisions, scored and plotted as predicted probability against observed outcome, answers that before any threshold ships.

Hosting is the other choice. Cloudflare introduced its Clef models on October 1 with open weights, deployed through Workers AI ^[14]. OpenAI's Decisions API is in public beta and returns typed answers to fixed questions ^[15]. Decision-1 starts in Foundry, and Microsoft says OpenRouter access is coming soon, without a date ^[16]. Thresholds tuned against one model's probabilities do not move to another model unchanged, so switching scorers means calibrating again. My context is a workflow already on Azure with a labeled eval set. There, I would trial Decision-1 first on support-ticket routing, one of the uses Microsoft names ^[18], because a misrouted ticket is cheap to reverse.