Microsoft-Decision-1, our model for fast decision-making

Microsoft has unveiled its new decision-scoring model, Microsoft-Decision-1, which outperforms both large language models (LLMs) and other decision models in terms of latency and quality. Decision models, unlike LLMs, are designed to deliver structured outputs that can be acted upon immediately by software, unlocking a wide range of useful tasks.

Microsoft-Decision-1 is available in Microsoft Foundry and through OpenRouter, enabling easy incorporation of decision intelligence into existing applications, agents, and workflows in a secure, trusted environment. The model excels in routing, classification, prioritization, verification, and workflow control, providing a calibrated probability score for each option and supporting various types of decisions, such as yes/no, multiple-choice, rating, and rubric-based grading.

Microsoft-Decision-1 achieved the highest accuracy in a 36-benchmark comparison, spanning nearly 150,000 questions across blind-trained benchmarks. It was also the fastest measured, being 2.5 times quicker than H2O-Lightning-4B v1.1, the runner-up, and 35 times faster than GPT-6 Sol. To ensure reliability, the model was evaluated across dozens of benchmarks, including routing, ranking, long context, multilingual, out-of-distribution, reasoning, and safety tasks, as well as comparisons with top-performing public models on the JevBench leaderboard.

Microsoft-Decision-1 demonstrated strong generalization and consistent performance across various tasks, even when presented with paraphrased options or reversed choices. The model successfully refused harmful behavior while maintaining a high degree of utility, successfully rejecting harmful requests in 5,250 requests across 11 benchmarks without needlessly blocking harmless ones.

Microsoft-Decision-1 has been successfully utilized in various applications, including XBOX Research processing feedback and reviews, the Copilot team measuring chat and agentic responses, on-call engineers retrieving relevant knowledge for live incident response, and Microsoft Discovery implementing an adaptive replanning feature for long-running scientific experiments.