Microsoft has released Microsoft-Decision-1, a model focused on scoring predefined choices in AI applications. It is designed for request routing, ticket classification, content filtering, and evaluating outputs and agent actions. Rather than generating a free-form answer, it returns a probability score for each option through an API.
The API scores every supplied option
Each request includes the task itself and a fixed set of possible answers. Developers can also supply grading criteria. The structured API response contains a probability score for every option, rather than just a single selected answer.
Supported tasks range from simple yes-or-no decisions to multiple-choice questions. The model also works with rating scales and rubrics that define grading rules. Evaluating AI responses and individual agent actions fits within the same framework.
Microsoft describes the returned probabilities as calibrated. According to the company, predictions assigned a score of 0.9 should be correct roughly nine times out of ten across representative cases receiving that score.
A single pass instead of text generation
Decision-1 is based on Qwen3.5-9B, which Microsoft post-trained to score options in one pass through the model. Its execution path excludes free-form text generation and repeated tool-calling loops, as well as multistep reasoning traces.
That design defines its scope: the model is not intended for free-form writing, open-ended planning, executing tools, or explaining its reasoning process. Every task must be expressed as predefined options or scores governed by a grading rubric.
Microsoft compared speed and decision stability
Microsoft says it evaluated Decision-1 on 36 benchmarks containing nearly 150,000 questions that were excluded from training. The company reports that the model achieved the highest accuracy and lowest latency among the systems included in that comparison.
At median (P50) latency, Microsoft's results show Decision-1 was 35 times faster than GPT-6 Sol and 4.5 times faster than Quyet-1.0-Large.
The company also examined how input changes affected decisions. Across eight types of input modification, it reports an average decision-flip rate of 1.3%. In tests that paraphrased option descriptions, reversed their order, or shuffled them randomly, Microsoft recorded no decision changes.
Deployments in Xbox Research, Copilot, and Microsoft Discovery
In an internal Xbox Research deployment, the model sorted more than 10,000 survey responses and reviews from Steam and X into predefined themes. Microsoft reports quality comparable to GPT-6 Sol at more than 14 times the speed for that task.
In Copilot, Decision-1 graded the quality of chat and agent responses. Microsoft says it delivered quality comparable to GPT5.6 Luna while running 100 times faster.
Another internal use involved responding to operational incidents. The model selected relevant information from logs, tickets, calls, messages, and other incident data. Microsoft reports better retrieval quality and greater speed than an unspecified baseline LLM.
In Microsoft Discovery, the model graded experiments for an agent that revises and repeats its plan. Microsoft reports 46 times more consistent scoring at three times the speed in this deployment. Adaptive replanning was nearly four times faster, according to the company.
Foundry availability and planned variants
Microsoft-Decision-1 is available in the Microsoft Foundry catalog. Pricing is $0.042 per million input tokens, with no charge for output tokens.
Microsoft also plans variants based on its own MAI models and on OpenAI models. For those future versions, the company intends to retain the integration pattern in which applications supply a fixed set of options.
