Microsoft has released Microsoft-Decision-1, a model specialized in structured decisions and classification, post-trained from Alibaba's Qwen3.5-9B. Available in Microsoft Foundry and OpenRouter, the model costs $0.042 per million input tokens with free output.
What happened?
Announced on October 9, 2026, Decision-1 does not generate free-form text. Instead, it receives a situation and a question with fixed options (yes/no, multiple choice, rating, rubric evaluation) and returns a calibrated probability for each option in a single pass.
According to Microsoft, the model achieved the highest average accuracy in a 36-benchmark comparison with nearly 150,000 questions kept blind from training. It was also the fastest: 4.5 times quicker than the runner-up and about 35 times faster than GPT-6 Sol in p50 latency.
Why does it matter?
Decision models are emerging as a distinct category from traditional LLMs. Instead of spending tokens generating text, they solve routing, verification, classification, and agent control tasks more cheaply and predictably. Companies like TypeSafe (with Jev), Cloudflare, and now Microsoft are investing in this approach.
Using a Chinese open-weight model (Qwen) as the base, with Microsoft post-training, illustrates how the industry is combining open models with proprietary optimizations for specific use cases.
What changes in practice?
Developers can use Decision-1 for tasks such as agent routing, intent classification, groundedness verification, data labeling, and workflow control — all at very low cost. Free output and low latency make it suitable for systems that need many decisions per second.
The model is available in public preview on Microsoft Foundry. Microsoft plans future rebases on its own MAI models and OpenAI's.
Image credit: Microsoft — Source: Microsoft Command Line
By GeekikiBot