Microsoft had a choice when it set out to build a model that answers with a probability instead of a sentence. It could have started from one of its own systems, or from a model built by OpenAI, its closest partner. It started instead with Qwen3.5-9B, an open-weight model from Alibaba Cloud, the cloud arm of the Chinese e-commerce group.
The result, announced this month on Microsoft’s Command Line blog by Achint Srivastava, a vice president of software engineering in the Office of the CTO, is called Microsoft-Decision-1. It belongs to a category the industry has begun to call decision models. Unlike the large language models that write prose and reason through open-ended problems, a decision model takes a fixed list of possible answers and assigns each one a calibrated probability. It does not compose a paragraph. It returns a score a piece of software can act on immediately.
The uses are deliberately unglamorous: routing a request to the right system, classifying a document, prioritizing a queue, verifying data, moderating content, checking code, grading the actions of an AI agent against a rubric. Microsoft says its new model supports yes-or-no questions, multiple-choice and rating scales, all through a single structured call.
The company’s benchmarks put it at the front of a crowded field. Microsoft says Decision-1 is 2.5 times faster than H2O-Lightning-4B and 2.8 times faster than Jev, a decision model released about three weeks earlier by the startup TypeSafe AI, in latency tests. It claims the highest accuracy, 83.5%, across 36 benchmarks, and second place on confidence scoring at 92.2%. On text classification, the company says, it costs more than 20 times less than OpenAI’s GPT-6 Sol.
For developers, the appeal is narrow and practical. A model that returns a score in a single pass can sit inside a workflow without adding a noticeable delay, and at that price it can run on every incoming request rather than only the ones a company has decided are important enough to pay for.
The pricing makes the pitch concrete. Microsoft lists the model at 0.042 dollars per million input tokens, with output tokens free — an unusual arrangement that reflects how decision models work. Because the model picks among options rather than generating text, its output is short by design, and the cost structure follows.
Srivastava framed the category in plain business terms. Decision models, he wrote, are purpose-built to deliver structured outputs that software can act on right away, and doing classification and decision work cheaply and quickly unlocks a range of practical tasks once a company understands the capability. Satya Nadella, Microsoft’s chief executive, said on social media that the company is already testing the model across its own operations, including incident response, quality control and scientific discovery.
The choice of Qwen as a foundation is the part likely to draw attention. Qwen is developed by Alibaba Cloud, and open-weight releases from Chinese labs have become a serious force in the market over the past two years, pressing prices down and forcing Western companies to answer with free releases of their own. Microsoft has not explained why it chose Qwen3.5-9B, though the economics of a nine-billion-parameter open model are easy to read: it is small, cheap to run and available without a negotiation.
A person close to the company said the choice reflected speed to market rather than any shift in strategy, and that the underlying model would change. Microsoft says the reliance is temporary. The company plans to rebuild Decision-1 on other models, including its own MAI family and systems from OpenAI, which would move the product onto foundations it controls. It is a familiar pattern for the company, which has built products on partners’ technology while developing substitutes in parallel.
The launch also says something about how the decision-model niche is filling up. TypeSafe AI’s Jev arrived first, and more than a hundred models now compete in the category, according to one count. The contest is not about who can reason most deeply. It is about who can return a trustworthy score fastest and cheapest, at a volume where a fraction of a cent per call becomes the deciding factor.
That is a different kind of competition from the race to build the largest model. It rewards small models, disciplined engineering and aggressive pricing rather than raw scale, and it plays to the strengths of companies that already sell software and cloud services to businesses. Microsoft’s willingness to lean on an Alibaba foundation, at least for now, suggests it wants to be early in that market more than it wants to be self-sufficient in it. The model is available through Microsoft Foundry and will soon appear on OpenRouter, putting it in front of developers who are already shopping for the lowest-cost way to make a routine decision.
