Microsoft launches Decision-1 for fast decision scoring

Microsoft says it created Decision-1 by post-training Qwen3.5-9B for single-pass decision scoring.[1] The model returns a calibrated probability for each option in a fixed answer set, and supports yes/no, multiple-choice and rating options.[1] Microsoft also says it can grade AI responses and agent actions against rubrics.[1] The company distinguishes decision models from language models designed to generate text or reason through complex problems: decision models produce structured outputs that software can act on immediately.[1] Microsoft designed this model for routing, classification, prioritization, verification and workflow control.[1]

Microsoft says Decision-1 had the highest accuracy in its comparison of 36 benchmarks, covering nearly 150,000 questions kept blind from training.[1] It also reported that the model was 2.5 times quicker than H2O-Lightning-4B v1.1 and 35 times quicker than GPT-6 Sol in its benchmarking.[1] These are Microsoft’s reported comparisons, not independent results: P.K. Sharma reported that independent results were not available on October 9, 2026.[6] Microsoft says the model changed its decision on 1.3% of perturbations on average across eight variations, with zero flips when option descriptions were paraphrased or options were reversed or shuffled.[1]

Microsoft’s announcement is dated October 9, 2026, and says the model was available through Microsoft Foundry and OpenRouter at launch.[1] Vercel said it also made Decision-1 available through AI Gateway on October 9, and that developers could call it through its AI SDK decision API, an OpenAI-compatible Decisions API or a TypeSafe-compatible API.[5] Vercel said an AI Gateway API key was sufficient, without a separate Azure account or deployment.[5] MarkTechPost reported a price of $0.042 per million input tokens, free output and a 32,768-token context window.[3]

Microsoft says the model can add decision intelligence to existing applications, agents and workflows in a secure, trusted environment.[1] It reported testing the model across the company for incident response, quality control and scientific discovery.[2] MarkTechPost reported that the model does not provide explanations or rationales, and that it is not intended for sole automated decisions involving credit, employment, housing, healthcare or legal rights.[3] The same report says applications should define options, thresholds and escalation paths, and provide human oversight.[3] Those limits matter when teams decide which workflows are suitable for automated scoring.