Microsoft releases Microsoft-Decision-1, a Qwen3.5-9B-based scoring model

Microsoft reports that the model achieved an average accuracy of 83.5% across 36 internal benchmarks, outperforming Quyet-1.0-Large at 81.9%. It also demonstrated a p50 latency of 85 ms, which Microsoft claims is 4.5 times faster than Quyet-1.0-Large and 35 times faster than GPT-6 Sol. However, the latency comparison is disputed by H2O.ai, which argues that Microsoft’s use of JevBench adjusted medians inflates competitor times. Independent verification on the JevBench board is still pending, where H2O-Lightning-4B currently ranks first.

Pricing is set at $0.042 per million input tokens, with output tokens free. The model is restricted to text-only inputs and cannot handle images, audio, or video. It is explicitly not designed for text generation, open-ended Q&A, or sole automated decisions in sensitive areas like credit or employment. Developers must define specific options and thresholds for use.