Microsoft introduces Decision-1 for fast AI workflow decisions

Microsoft has introduced Microsoft-Decision-1, a model designed to make quick, structured choices for AI applications and automated workflows. In an official Microsoft blog post, Achint Srivastava announces the model and describes it as a tool for tasks such as classification, model routing, quality checks, and safety screening.

Unlike general-purpose large language models, Decision-1 is meant to select from predefined options rather than generate lengthy text. Developers can provide yes or no choices, multiple-choice answers, ratings, or assessment criteria. The model returns a probability for each option, which can help software decide whether to act automatically, ask for human review, or pass a task to another system.

It’s a bigger trend started by TypeSafe’s Jev. Read more about Decision Models here.

Microsoft reports speed and accuracy gains

Microsoft says it post-trained Qwen3.5-9B to produce decisions in a single pass. According to the company’s own tests, Decision-1 achieved the highest accuracy across 36 benchmarks containing almost 150,000 questions that were withheld from training. Microsoft also says its model ran 2.5 times faster than H2O-Lightning-4B v1.1 and about 35 times faster than GPT-6 Sol in its latency measurements.

The company also tested whether small changes in wording or formatting altered results. It reports that the model changed its decision in 1.3 percent of such tests on average. Microsoft says it found no decision changes when options were reordered or their descriptions were paraphrased.

Internal teams are already using the model for tasks including sorting Xbox customer feedback, grading Copilot responses, incident-response routing, and evaluating scientific experiments. Microsoft says Xbox Research found the model competitive with GPT-6 Sol for feedback labeling while operating faster and at lower cost.

Microsoft-Decision-1 is available through Microsoft Foundry and OpenRouter. Microsoft lists input pricing at $0.042 per million tokens, while output tokens are free.