Microsoft-Decision-1: Microsoft launches a model that does not generate text but makes decisions, 35 times faster than GPT-6 Sol

Derived from Qwen3.5-9B and available in Foundry and on OpenRouter, the model assigns a calibrated probability to each option for routing, classifying or verifying within agent workflows.

Microsoft presents Microsoft-Decision-1, a "decision model" offered in Microsoft Foundry and via OpenRouter. Unlike an LLM, it does not write anything: faced with a fixed set of possible answers (yes/no, multiple choice, score, evaluation grid), it returns a calibrated probability for each one, which can be used directly by software. The intended uses are the routing of requests, the classification, the prioritization, the verification and the control of an agent's steps, including the scoring of responses or actions of other models.

Latency as the main argument

Microsoft obtained it by post-training Qwen3.5-9B and plans to rebuild it soon on other bases, including its own MAI models and those of OpenAI. The publisher insists on the cumulative cost of chained decisions: 100 milliseconds more over 20 successive stages, that's two seconds added to the course. According to its measurements, the model is 2.5 times faster than its immediate follow-up, H2O-Lightning-4B v1.1, and about 35 times faster than GPT-6 Sol in median latency.

In terms of quality, Microsoft claims the best accuracy in a comparison of 36 benchmarks, i.e. nearly 150,000 questions that are kept out of training. It also highlights stability: by reformulating or reordering the same query in eight ways, the decision changes in only 1.3% of cases on average. These figures come from the publisher itself and remain to be confirmed by independent evaluations.