微软推出Microsoft-Decision-1决策模型 性能和成本均处业界领先

微软推出了专门用于快速决策任务的AI模型Microsoft-Decision-11。这一模型在36个基准测试中表现突出，平均准确率达到83.5%，排名第一1。该模型在速度方面优势明显，运行速度比GPT-6 Sol快35倍，同时成本仅为其1/2001。

Microsoft-Decision-1针对路由、分类和优先级排序等结构化决策任务进行了专门设计，不同于通用的大语言模型1。模型的中值延迟为85毫秒（p50），p95延迟为125毫秒1。在校准分数上，该模型获得92.2分，仅次于Quyet-1.0-Large的93.1分1。模型在鲁棒性测试中表现稳定，在8种扰动条件下仅有1.3%的决策出现变化1。

该模型已在微软内部多个场景中应用取得成效。Xbox研究团队利用其处理了超过10,000条反馈，处理速度比GPT-6 Sol快14倍1。微软的Discovery自适应复制规划应用通过部署该模型将速度提升了4倍1。Microsoft-Decision-1基于Qwen3.5-9B进行后训练，在Microsoft Foundry平台和OpenRouter上提供1，输入令牌定价为$0.042/百万，输出令牌免费1。

Microsoft has unveiled Microsoft-Decision-1, a specialized artificial intelligence model engineered for rapid, cost-effective execution of structured decision tasks 1. Unlike large language models, this model targets specific decision-making functions such as routing, classification, and prioritization, and is now available on the Microsoft Foundry platform and OpenRouter 1.

The model demonstrates commanding performance across multiple dimensions. Microsoft-Decision-1 achieved the highest average accuracy of 83.5% across 36 benchmarks encompassing 147,137 questions 1. It operates at exceptional speed, running 4.5 times faster than Quyet-1.0-Large and 35 times faster than GPT-6 Sol, with median latency of 85 milliseconds and p95 latency of 125 milliseconds 1. Economically, the model offers dramatic cost advantages—input tokens are priced at $0.042 per million with free output tokens, making it approximately 200 times cheaper than GPT-6 Sol 1. The model is built on post-trained Qwen3.5-9B infrastructure 1.

Beyond benchmark metrics, Microsoft-Decision-1 demonstrates robust real-world performance. Its calibration score of 92.2 ranks second, trailing only Quyet-1.0-Large's 93.1 1, and it maintains high consistency under adversarial conditions, with only 1.3% of decisions changing across eight perturbation tests 1. Internal applications have already yielded substantial benefits: Xbox research deployed the model to process over 10,000 feedback entries 14 times faster than GPT-6 Sol, while Microsoft Discovery's adaptive replication planning achieved a fourfold speed improvement 1.