Microsoft launches Decision-1, a fast model for routing, classifying and scoring choices

Microsoft has introduced Microsoft-Decision-1, a new AI model built for fast decision-scoring. It is available in Microsoft Foundry and through OpenRouter.

Microsoft calls this a "decision model." Unlike a large language model (LLM), which writes text or works through complex problems, a decision model gives structured answers that software can act on right away. Microsoft says it is meant for routing, classification, prioritization, verification and workflow control.

Under the hood, the company post-trained Qwen3.5-9B for fast, single-pass scoring. It says it will soon rebase the model on others, including Microsoft AI (MAI) and OpenAI models. Given a fixed set of answer options, the model returns a calibrated probability score for each one. Calibrated means a 90% score should be right about nine times in 10. It handles yes/no, multiple-choice and rating options, and can grade AI responses and agent actions against a rubric, all through a structured API call.

On quality, Microsoft says the model had the highest accuracy in a 36-benchmark comparison covering nearly 150,000 questions, using benchmarks kept blind from training. On speed, it says the model was the fastest measured. It was 4.5 times quicker than Quyet-1.0-Large, the runner-up, and 35 times quicker than GPT-6 Sol.

Speed matters because decisions often run one after another. Microsoft gives an example: adding 100 milliseconds to each of 20 sequential decisions adds two seconds to the whole workflow.

Microsoft also reports results on robustness. It changed the same request in eight ways, such as paraphrasing, reordering options and adding formatting noise, and measured how often the decision flipped. The model changed its decision on 1.3% of perturbations on average, with no flips when option descriptions were paraphrased or when options were reversed or shuffled.

On safety, the company tested the model on 5,250 requests across 11 benchmarks covering harmful content, jailbreak attempts and prompt injection. It says the model refused harmful behavior while staying useful for harmless requests.

Microsoft also shared internal tests. Xbox Research used the model to sort more than 10,000 pieces of feedback and reviews from surveys, Steam and Twitter/X into themes set by researchers. They found it competitive with GPT-6 Sol on quality, while running over 14 times faster and 200 times less expensive.

The Copilot team used it to measure the quality of chat and agentic responses. They found it competitive with GPT5.6 Luna and 100 times faster. Microsoft's on-call engineers found it performed better and faster than an LLM at retrieving knowledge during live incidents.

In Microsoft Discovery, an agent reviews an earlier experiment, revises its approach using rubric grades and repeats. Microsoft says the model scored 46 times more consistent than an LLM-based score at three times the speed, and adaptive replanning ran nearly four times faster.

Microsoft lists many other possible uses. They include controlling agents, picking the best model for a request, labeling data, judging AI answers, matching user intent, validating data, scanning code, screening for safety risks, and choosing actions in computer or robotics tasks.

Pricing is $0.042 per million input tokens. Output tokens are free. Developers can start in Microsoft Foundry or through OpenRouter.