Microsoft has introduced Microsoft-Decision-1, a decision-scoring model based on Qwen3.5-9B. Unlike standard generative AI, this model evaluates a fixed set of options and returns probability scores in JSON format. It is designed for direct integration with software systems.
The model was trained using public datasets and synthetic data. In testing across 36 benchmarks with 147,137 questions, it achieved an average accuracy of 83.5%. Microsoft reports a p50 latency of 85 ms and p95 latency of 125 ms, noting these speeds outperform several competitors.
Robustness testing involved 5,250 requests, showing the model maintained consistency across various perturbations. Microsoft claims high utility in safety tests, including resistance to jailbreaks and prompt injections. The model is currently available via Microsoft Foundry and the OpenRouter Decisions API.
Pricing for the service is set at $0.042 per million input tokens, with no charge for output tokens. This compares to the $0.10 per million input tokens charged by OpenAI's Luna decisions endpoint. The model is not compatible with standard chat completion SDKs.
