Microsoft Releases Decision-1, a Model That Only Makes Decisions
Microsoft has released Microsoft-Decision-1, a model that takes a request plus a fixed set of options and returns a calibrated probability for each in a single pass: yes/no, list selection, ranking, or rubric-based scoring of an AI answer or agent action. It is built on Qwen3.5-9B after post-training, and Microsoft reports the best accuracy among both LLMs and other decision models across 36 closed benchmarks and nearly 150,000 questions, while being 4.5x faster than the nearest competitor and 35x faster than GPT-6 Sol. Decisions change in only 1.3% of cases when options are paraphrased or shuffled.
Microsoft says output tokens are free and input costs $0.042 per million tokens, and the model is already in internal use: Xbox Research labeled 10,000 reviews at GPT-6 Sol quality but 14x faster and 200x cheaper, while Copilot evaluates agent responses 100x faster than GPT-5.6 Luna. Decision-1 is available now in Microsoft Foundry, with an OpenRouter release planned.
Related: OpenAI’s Decisions API Returns a Choice, Not a Paragraph, TypeSafe AI Introduces System One Models and Jev, Microsoft Releases Two New MAI Models for Image Generation and Speech