Microsoft released Microsoft-Decision-1 on October 9, 2026, a decision-scoring model rather than a text generator. Given a fixed set of answer options, such as yes or no, a multiple choice or a rating, it returns a probability for each option in a single pass. Microsoft targets routing, classification, prioritization, verification and workflow control, plus rubric-based grading of AI responses and agent actions. It is available in Microsoft Foundry and through OpenRouter at $0.042 per million input tokens, with output tokens free.
The model is Qwen3.5-9B post-trained for decision scoring, and Microsoft says it will rebase it on other models, including its own MAI models and OpenAI’s. Microsoft reports the highest accuracy in a 36-benchmark comparison covering nearly 150,000 questions held out from training, and a median latency about 35 times faster than GPT-6 Sol. It says the model changed its decision on 1.3 percent of input perturbations on average, with no flips when options were paraphrased, reversed or shuffled. In internal use, Microsoft says it matched GPT-6 Sol on Xbox feedback labeling at over 14 times the speed and 200 times lower cost.
Why it matters: much of an agent pipeline is small yes-or-no decisions, which routers and judges have been handling by running a full generative model. A dedicated, calibrated scorer is a cheaper building block, and Microsoft’s release came eight days after Cloudflare’s Clef decision models, a sign that decision models are becoming their own category.
What it does not show: all comparisons are Microsoft’s, the calibration claim comes with no calibration measurement, and the approach only applies where the answer options are known in advance.