What shipped
Achint Srivastava, vice president of software engineering in Microsoft’s Office of the CTO, introduces Microsoft-Decision-1 in a post dated 2026.10.09. The post calls it a model for fast decision-scoring. It is available in Microsoft Foundry and through OpenRouter. It is designed for routing, classification, prioritization, verification, and workflow control, so teams can put it into existing applications, agents, and workflows.
The post separates decision models from language models. Language models, it says, generate text or reason through complex problems. Decision models are purpose-built to deliver structured outputs that software can immediately act on. That is Microsoft’s description of the category, not an independent taxonomy. The post also calls decision models an important category now emerging in AI. That judgment comes from the same product post.
[1]
Microsoft’s own comparison
The post says Microsoft-Decision-1 achieved the highest accuracy in Microsoft’s 36-benchmark comparison. Those benchmarks span nearly 150,000 questions and were kept blind from training. In the same measurements it was the fastest: 4.5 times quicker than Quyet-1.0-Large, the runner-up, and 35 times quicker than GPT-6 Sol. The post also says that on structured decision tasks its latency and quality beat language models and other decision models. The deck and the body both use that comparison.
The question count, the comparison targets, and the speed ratios all come from Microsoft’s own test. The post does not name the 36 benchmarks, and it does not include a reproduction from outside Microsoft. In the same passage the author writes that decisions and classification can be done at very low cost with high performance. The post gives no unit price, so that sentence stays a product description, not a published price. A reader checking the ranking can see only two named systems: Quyet-1.0-Large in second place, and GPT-6 Sol at the other end. The post does not describe how either system was configured.
[1]What the post does not say
The official post does not give a price, does not say whether the weights are public, and does not give a context length. This article does not treat “better than language models” as a result already checked by an outside group. Quyet-1.0-Large and GPT-6 Sol appear only as the comparison targets Microsoft named. What a reader can check today is where the model is available, the five designed uses, and the self-reported figures above.
[1]