Microsoft has launched Microsoft-Decision-1, a specialized model designed for data classification, routing, prioritization, and workflow management. The system is built upon the open-source Qwen3.5-9B model developed by Alibaba and is optimized for structured decision-making rather than text generation.
The model assigns calibrated probabilities to fixed sets of options, supporting binary choices, multiple-choice selection, and ranking tasks. Microsoft reports that the model outperformed competitors across 36 benchmarks, processing nearly 150,000 unseen questions with high accuracy. It is noted for its speed, performing 2.5 times faster than H2O-Lightning-4B v1.1 and 35 times faster than GPT-6 Sol.
In practical applications, the Xbox Research division utilized the model to categorize over 10,000 user reviews, achieving performance comparable to GPT-6 Sol while operating 14 times faster and at 200 times lower cost. The Copilot team also employed the model for quality assessment, matching GPT-5.6 Luna performance with a 100-fold increase in speed.
The model demonstrates high stability, with results changing in only 1.3% of cases during robustness testing. It also passed security evaluations involving 5,250 malicious prompts and jailbreak attempts. Microsoft plans to integrate the decision-making architecture with other base models, including its own MAI and OpenAI developments.
Microsoft-Decision-1 is currently available at a price of $0.042 per million input tokens, with no charge for output tokens. Potential use cases for the technology include content moderation, code verification, robotics, and adaptive experiment planning.
