Microsoft has released Decision-1, a small model built for classification, evaluation and routing rather than open-ended text generation. It is available through Microsoft Foundry and OpenRouter, where input costs $0.042 per million tokens and output tokens are free.
The model is based on Qwen3.5-9B. Instead of producing an essay, it can return a probability, select from a defined list or assign a score. That format suits agent control tasks such as choosing which tool or larger model should handle a request, while avoiding the latency and expense of generating unnecessary prose.
Microsoft reports 83.5 percent accuracy with 85-millisecond latency across 36 benchmarks containing nearly 150,000 questions. It also says Decision-1 was 2.5 times faster than the next model in its comparison, H2O-Lightning-4B. These are vendor results, and Cloudflare’s recently released Clef decision models were not included, so independent comparisons remain necessary.
Decision-1 joins a quickly growing category that includes products from Jev, OpenAI and Cloudflare. Its low listed price makes large-volume routing plausible, but benchmark averages do not show performance on every organization’s labels or risk thresholds. Developers should test calibration, failure modes and fallback rules on their own data before allowing a fast classifier to make consequential choices for an agent.