The model does not generate text. It takes a fixed set of options and returns a calibrated probability for each through a structured API call, and it covers yes/no, multiple-choice, rating and rubric-based grading. Microsoft post-trained Alibaba's open-weight Qwen3.5-9B, which runs one pass over up to 32,768 tokens and emits zero output tokens. The weights are not distributed. Reports disagree on its status. Microsoft's blog calls it public preview, while other coverage lists it as generally available since October 8. The launch is led by Achint Srivastava, a Microsoft vice president who previously co-founded the AI evaluation startup Pi Labs. OpenAI's Luna decisions endpoint charges $0.10 per million input tokens. Microsoft claims the top accuracy across 36 benchmarks covering nearly 150,000 questions withheld from training, but the catalogue page's Benchmarks tab contains a methodology paragraph and no figures. Routing, judging and guardrail checks are now a cheap, hosted model call. Startups that sell evaluation or routing tools will now be compared against Microsoft's price. Teams building agents should test the model on their own data first, because the published accuracy claims come only from Microsoft.