Microsoft-Decision-1 is designed for structured AI decisions

Microsoft describes Decision-1 as a purpose-built decision model whose structured outputs can be acted on directly by software.[1] The company designed it for routing, classification, prioritization, verification and workflow control across applications, agents and workflows.[1] For a fixed set of answer options, the model returns a calibrated probability score for each option.[1] It supports yes-or-no, multiple-choice and rating decisions, along with rubric-based grading of AI responses and agent actions.[1] This focus differs from a request to produce a written explanation: Microsoft’s documentation says Decision-1 does not generate a written rationale.[2]

Microsoft says it built the model by post-training Qwen3.5-9B for fast, single-pass decision scoring, and plans to rebase it on other models, including models from Microsoft AI and OpenAI.[1] In its comparison of 36 benchmarks covering nearly 150,000 questions withheld from training, Microsoft reports that Decision-1 had the highest accuracy.[1] The company says it ran 2.5 times faster than the runner-up, H2O-Lightning-4B v1.1, and 35 times faster than GPT-6 Sol in its benchmarking.[1] InfoWorld reports that the benchmark results and cost calculations have not been independently verified.[3] Treat the performance figures as Microsoft’s reported results, not an independent assessment.

Microsoft says Xbox Research used Decision-1 to categorize more than 10,000 open-ended feedback items and reviews drawn from surveys, Steam and Twitter/X.[1] Microsoft reports that the model was competitive in quality with GPT-6 Sol for that work, while running more than 14 times faster and costing 200 times less.[1] The company lists input-token pricing at $0.042 per million tokens and says output tokens are free.[1] The example offers teams a concrete task to evaluate when considering the model, but it does not establish what savings another organization would achieve. Kanerika AI development manager Amit Kumar Jena cautioned organizations against using Microsoft’s benchmark results to predict their enterprise savings.[3]

Microsoft says it tested Decision-1 on 5,250 requests across 11 safety benchmarks covering harmful content, jailbreak attempts and prompt injection, and that the model refused harmful behavior while retaining a high degree of utility.[1] Microsoft also reports that the model changed its decision on an average of 1.3% of tested input perturbations, with no flips in tests that paraphrased answer descriptions or reversed or shuffled options.[1] Its documentation warns that scores can change with the wording or ordering of questions and options, and may reflect bias in the base model and training data.[2] Forrester analyst Charlie Dai said the lack of reasoning can make decisions harder to explain, validate and challenge, and recommended human review for high-impact decisions.[3]