Microsoft introduced Microsoft-Decision-1 on October 9, 2026 as a model for structured decision scoring. The company says it is available through Microsoft Foundry and OpenRouter. That is a product announcement, not independent proof that it is the best choice for every workflow.
The first version is post-trained from Qwen3.5-9B. Microsoft says it plans to rebase future versions on other models, including Microsoft AI and OpenAI. A planned change should not be described as the architecture of the model available today.
The model takes a fixed set of answer options and returns a probability score for each. Microsoft lists yes/no, multiple-choice and rating tasks, plus rubric-based grading of AI responses and agent actions. Scoring predefined options is different from generating an unrestricted prose answer.
Microsoft reports the highest accuracy in its comparison across 36 benchmarks and nearly 150,000 questions kept blind from training. It also reports being 4.5 times faster than Quyet-1.0-Large and 35 times faster than GPT-6 Sol in its benchmarking. These are company-run comparisons, not independent measurements or guarantees for a different workload.
The announcement lists input pricing at $0.042 per million tokens, with output tokens free. It also reports an average decision change on 1.3 percent of input perturbations. Neither a low token price nor a calibrated score removes the need to test representative cases and decide when human review is required.
Limits: the source does not establish zero errors, universal latency or a complete cost for deploying an application. Availability, price and comparative results are reported as stated in the dated announcement. A decision model can still choose the wrong option, so its score should not be treated as permission to take a consequential action.