Microsoft Decision Model Debuts: A New Era for AI-Assisted Choices
In a recent announcement on its official blog, Microsoft introduced a new AI model designed specifically for decision-making: Microsoft-Decision-1. Now available on Microsoft Foundry and OpenRouter, this model targets scenarios that require clear-cut choices, such as routing, classification, prioritization, validation, and workflow control.
Benchmark Dominance: Unmatched Accuracy and Speed
The model's performance metrics are compelling. In a broad set of 36 benchmark tests covering nearly 150,000 questions—all conducted under blind testing conditions during training—Microsoft-Decision-1 achieved the highest overall accuracy rate.
Perhaps more critical for real-time applications is its inference speed. In practical tests, it significantly outpaced competitors, processing requests 2.5 times faster than the second-place H2O-Lightning-4B v1.1 model and a staggering 35 times faster than GPT-6 Sol. This speed advantage could translate to major efficiency gains in automated systems.
How It Works: The Power of Calibrated Scores
Built through further training on the Qwen3.5-9B model, Microsoft-Decision-1 is slated for future integration into the Microsoft AI (MAI) stack and migration to OpenAI model architectures.
Its core function is straightforward: when presented with a fixed set of options, the model doesn't just pick one. Instead, it outputs a calibrated probability score for each potential choice. This provides a nuanced view of the AI's "reasoning," allowing downstream systems to make more informed use of its output.
Built for the Real World: Robustness and Cost
Decision consistency is key. Microsoft tested the model's robustness by applying eight different text perturbations (like rephrasing or adding noise) to the same requests. On average, only 1.3% of cases resulted in a decision "flip" (changing from option A to B), indicating high stability against input variations.
Pricing is structured to be attractive for high-volume use: $0.042 per million tokens for input, with output being completely free. This model is particularly cost-effective for applications that involve analyzing large amounts of text to arrive at a concise decision.
Proven in Practice: An Internal Use Case
The model has already moved beyond testing. Internally, Microsoft's Xbox Research team deployed Microsoft-Decision-1 to process and categorize over 10,000 pieces of user feedback.
The results were striking: the categorization quality matched that of the powerful GPT-6 Sol model, but the task was completed 14 times faster and at a computational cost 200 times lower. This real-world application highlights its potential for delivering both quality and dramatic efficiency improvements.
As businesses seek smarter, faster, and more economical tools for automating complex decision loops, Microsoft-Decision-1 enters the field as a potent new contender designed to meet that demand.