Microsoft Releases 'Microsoft-Decision-1', a Jev-Compatible Decision Model

Microsoft officially launched (GA) the Jev-compatible API format decision model 'Microsoft-Decision-1' on October 9, 2026 (October 8 per the model card). The model was developed by post-training Qwen3.5-9B for single-pass decision scoring, characterized by returning only probabilities per option without generating sentences. It is optimized for classification, routing, prioritization, verification, and workflow control.

Pricing and Specifications

Input costs $0.042 per 1M tokens, while output is free. The context length is 32,768 tokens, supporting text input only. Question types consist of three kinds: noul (true/false), choice (multiple choice), and score (ordinal scale). It is available via Microsoft Foundry and OpenRouter (Azure provider).

Performance and Benchmarks

In Microsoft's internal benchmark (147,137 questions), it ranked first with an average accuracy of 83.5%. However, on the independent benchmark JevBench (1,500 questions), it ranked 7th with a Capability score of 70.8, lower than Jev (77.1) and OpenAI Decisions (73.5). A gap in calibration scores compared to Jev also exists. Latency was reported as p50 85ms (Microsoft measurement), but appeared as p50 0.5s in JevBench measurements.

Considerations for Adoption

The performance advantage is based on a private test set selected by Microsoft, which differs from independent measurement results. In particular, Korean performance is 33.1 points on JevBench, lower than English (56.8). Performance degradation is expected in low-resource languages and specialized domains (medical, legal, etc.), so accuracy and calibration should be pre-validated with actual business data.