Microsoft Decision-1: 9B Qwen Model Launch Explained
Short summary
Microsoft released a narrow model that doesn't generate text. It scores a fixed set of answer options and returns a confidence for each, aimed at routing, classification, and verification in production pipelines. It's built on Alibaba's open-weight Qwen3.5-9B, and most of the performance numbers so far come from Microsoft itself.
- The model returns calibrated probabilities over predefined options in a single forward pass, so applications can act on a score instead of parsing free-form output.
- Microsoft says it beat 36 benchmarks covering about 150,000 questions and runs far faster than GPT-6 Sol on P50 latency, but these are vendor-reported figures with no independent reproduction yet.
- Pricing reported via OpenRouter is $0.042 per million input tokens with free output, and weights aren't released, so access is through hosted APIs only.
Full summary
Microsoft's Microsoft-Decision-1, launched October 9, 2026, is deliberately narrow. It doesn't chat, summarize, or write. Given content and a fixed set of answer options, it scores each option in a single forward pass and returns a calibrated probability for each. Microsoft pitches it for routing, classification, prioritization, verification, and workflow control: sending support tickets to the right team, making yes/no moderation calls, or checking whether an AI agent's output should be allowed to proceed. It's built on Alibaba's open-weight Qwen3.5-9B, a 9-billion-parameter model, and it's available through Microsoft Foundry and OpenRouter. The company has said it will eventually rebase the model on its own MAI models and OpenAI's, which makes the Qwen base look like a speed-to-market choice rather than a long-term commitment.
The performance claims are large and specific, and they come entirely from Microsoft. The company says Decision-1 posted the highest accuracy across 36 benchmarks and nearly 150,000 questions, with secondary reporting putting average accuracy at 83.5 percent. Speed claims include roughly 35 times lower P50 latency than GPT-6 Sol and 2.5 times faster than H2O-Lightning-4B v1.1, with a reported response time near 85 milliseconds, though the hardware and test conditions behind that figure aren't spelled out. Note that the speed comparison pits a scoring model against a text generator, so the gap says less than it appears to. None of the coverage reviewed includes independent reproduction. The one number that may matter most in production is a claimed 98.7 percent perturbation stability, meaning decisions rarely flip under small input rewording. If that holds up outside Microsoft's testing, it would count for more than the raw speed figures.
Pricing, as far as anyone has reported it, is $0.042 per million input tokens with output free, according to OpenRouter's listing. Microsoft hasn't confirmed that number in its own materials. If it holds, the model is cheap enough to replace general-purpose LLMs on high-volume, low-complexity calls like spam filtering or ticket triage, where paying for a full model to answer a yes-or-no question has been hard to justify. The practical question for engineering teams is whether the stability and accuracy claims survive their own data, which is the test worth running before building anything on it.