Microsoft launches Decision-1, a 9B decision-scoring model for routing and verifying agent actions at $0.042 per million input tokens
Post-trained from Qwen3.5-9B, it returns a calibrated probability for each of a fixed set of options instead of generating text. Output tokens are free. Microsoft's speed and accuracy claims come from its own benchmarks.
Microsoft introduced Microsoft-Decision-1 on October 9, a model built for routing, classification, prioritization, verification and workflow control, available in Microsoft Foundry and through OpenRouter. Rather than generating an answer, it takes a fixed set of options and returns a calibrated probability for each. It handles yes/no, multiple-choice and rating questions, plus rubric-based grading of AI responses and agent actions. Pricing is $0.042 per million input tokens, and output tokens are free.
What's under it, and what Microsoft claims
Microsoft says it post-trained Alibaba's Qwen3.5-9B for single-pass decision scoring, and plans to rebase the model on Microsoft AI (MAI) and OpenAI models. The performance figures are all Microsoft's own. It reports the highest accuracy in its 36-benchmark comparison and the fastest speed it measured: 2.5 times quicker than the runner-up, H2O-Lightning-4B v1.1, and 35 times quicker than GPT-6 Sol. In one internal Xbox research labeling workload, it says Decision-1 was competitive with GPT-6 Sol on quality while running more than 14 times faster at 200 times lower cost. The probabilities are meant to be used directly: Microsoft's target is that a 90% prediction is right about nine times out of ten, so an application can decide when to act, defer or send a case for review.
Why it matters
A lot of agent spend goes to frontier models making small, structured calls: which tool to use next, whether an output passes, whether a human should see it. Microsoft is betting that work belongs on a cheap, specialized model, and its post says it plainly: "cost plays a major role in how people decide to use AI." The numbers are vendor benchmarks, so test the model on your own routing and grading traffic before trusting them. Still, if every classify-or-verify step in your agent pipeline runs through a frontier model, that line item is worth measuring.