Microsoft’s newest AI product does not write. Give it a list of options and it returns a probability for each one — a score, not a sentence. The company is betting that much of what AI does inside a business is decision, not prose, and that decisions can be priced cheaply enough to run at high volume.
The model, called Microsoft-Decision-1, arrived on October 9 in Microsoft Foundry, with access through OpenRouter. It reads the supplied content and returns a calibrated probability for each fixed answer option, so the response carries its own confidence and can tell an application when to act, when to defer, and when to ask a human to review.
The uses Microsoft lists are unglamorous and everywhere: classification, routing, prioritization, verification, workflow control, guardrails for AI agents, and judging the output of other models. The interface supports three shapes of question — a true-or-false judgment, a single choice among candidates, and a graded score against criteria.
Under the hood, the model is small. It is built on a nine-billion-parameter base and tuned for single-pass scoring rather than generation. Small and single-purpose is the point: the jobs it targets repeat thousands of times an hour, and each one has to come back fast and cheap.
The performance claims are specific. Microsoft says Decision-1 posted the highest accuracy in a comparison across 36 benchmarks spanning nearly 150,000 questions, all kept out of its training data. The company also puts its median response time at roughly 85 milliseconds, which it describes as about 35 times faster than its GPT-6 Sol model on the same task.
The price is the headline. Input costs 0.042 dollars per million tokens, and output is free. That is almost exactly the rate TypeSafe, the San Francisco startup behind the decision model Jev, published when it announced its own product in late September. The category is young, and the pricing is already converging.
Satya Nadella announced the model himself, and Microsoft said it uses Decision-1 internally for incident response and quality control. The internal adoption is a signal: a company does not run a model on live operations unless it trusts the scores enough to act on them.
Decision-1 lands in a market that barely existed two months ago. TypeSafe’s Jev drew an eight-hundred-seventy-million-dollar round from Andreessen Horowitz and a valuation of seven and a half billion dollars within weeks of launch. Cloudflare has been shipping its own Clef family of open decision models. The shared premise is that generating text is wasted work when the answer is a selection.
That premise matters because of cost. A generative model must be paid to produce every word of an answer even when the answer is “yes.” A decision model returns a number. At fractions of a cent per thousand inputs, the economics invite companies to move routing, filtering and scoring off expensive chatbots and onto models that only rank.
The shift has a consequence for the AI business. The companies that built large text models charge by the token for language. If a meaningful share of inference moves to scoring, the value accrues to whoever is cheapest and most accurate on a narrow, repeatable task, and the moat of a general-purpose model grows thinner at the edges.
For Microsoft, the product fits a familiar playbook. The model runs on Azure and slots into the Foundry catalogue alongside the generative systems, giving the company a second lane into the same customers. A developer who routes a customer-service queue with Decision-1 is still paying Microsoft, just for a smaller, faster model.
The product also answers a question Microsoft has been asking about its own cloud. Enterprises that adopted generative AI for chat now want the same systems to make routine calls without paying a full language model for each one. A scoring model gives Azure a cheaper tier to offer those customers, and a reason to keep the workload inside Microsoft rather than shipping it to a rival whose pricing is equally low.
Microsoft has also said the model’s weights will be updated continually while the interface stays fixed, a nod to how fast the decision-model field is moving. Buyers who build on it today are betting that accuracy improves without their integration work being redone.
For now, the category is defined by its claims. Blind-benchmark accuracy and median latency are the numbers every vendor will cite, and independent buyers will have to verify them against their own workloads before trusting a machine to route a customer or judge an answer.
The unresolved question is whether decision scoring becomes a commodity. The pricing already looks like one. If it does, the interesting contest is less between Microsoft and its rivals over accuracy than over who captures the high-volume, low-margin decisions that keep a business running.
