A new category of AI models is emerging to tackle complex decision-making tasks. Unlike large language models (LLMs), which generate text or reason through problems, these decision models are designed to deliver structured outputs that software can immediately act on.
These models have the potential to unlock a wide range of useful tasks by making decisions and classifying things at very low cost with high performance.
Microsoft has introduced its latest model in this category: Microsoft-Decision-1. This new AI tool is specifically designed for fast decision-scoring, available through Microsoft Foundry and OpenRouter.
The model’s primary function is to provide a calibrated probability score for each option given a fixed set of answer choices. It supports yes/no, multiple-choice, rating options, as well as rubric-based grading of AI responses and agent actions via a simple structured API call.
Microsoft-Decision-1 has achieved top performance in latency and quality on structured decision tasks, outperforming both LLMs and other decision models.
The model’s accuracy was tested across 36 benchmarks spanning nearly 150,000 questions. In this comparison, Microsoft-Decision-1 demonstrated the highest accuracy, with a significant lead over its competitors.
Quyet-1.0-Large, another prominent AI model, came in second place but still lagged behind Microsoft-Decision-1 by a considerable margin. GPT-6 Sol trailed far behind, taking 35 times longer to complete tasks compared to the new Microsoft model.
To build Microsoft-Decision-1, researchers post-trained Qwen3.5-9B for fast, single-pass decision scoring and plan to rebase it on other models in the future, including Microsoft AI (MAI) and OpenAI.
The development of reliable decision models requires addressing several key challenges: speed, quality that generalizes across tasks, robustness against input variations, probability and confidence calibration, and safety.
Speed is a critical issue as each decision adds delay to workflows. For example, adding 100 milliseconds to each of 20 sequential decisions can add two seconds to the overall workflow.
To address this challenge, Microsoft-Decision-1 boasts an impressive P50 latency of ~35x faster than GPT-6 Sol.
Quality that generalizes is another crucial aspect of decision models. It’s easy to overfit a model for one benchmark or type of task but essential to know whether the quality carries over to various tasks not included in training data.
To evaluate Microsoft-Decision-1, researchers tested it across dozens of benchmarks kept blinded from training, spanning routing, ranking, long context, multilingual and out-of-distribution tasks, reasoning, and safety. The model performed exceptionally well on these broader sets of benchmarks, demonstrating strong generalization.
Robustness is also a vital aspect of decision models. Equivalent inputs should produce equivalent decisions without being affected by paraphrased states or instructions, option descriptions changes, choices reordering, keys change, or harmless formatting noise.
The researchers perturbed the same request in eight ways and measured how often the decision flipped. Microsoft-Decision-1 showed remarkable stability with only 1.3% of perturbations causing a flip on average, including zero flips when option descriptions were paraphrased or options reversed/shuffled.
Probability and confidence calibration are essential for applications to decide when to act, defer, or ask for review based on the model’s predictions.
Safety is another critical aspect as decision models should recognize harmful requests without needlessly blocking harmless ones. Microsoft-Decision-1 was tested across 5,250 requests covering harmful content, jailbreak attempts, and prompt injection.
The results showed that the model successfully refused harmful behavior while retaining a high degree of utility.
Microsoft has been testing Microsoft-Decision-1 internally in various scenarios to evaluate its performance and potential applications. One such example is data labeling where XBOX Research used the model to process over 10,000 open-ended pieces of feedback and reviews from surveys, STEAM, Twitter/X, and sort them into a fixed set of themes established by researchers.
The results showed that Microsoft-Decision-1 was competitive on quality with GPT-6 Sol while running over 14 times faster and 200 times less expensive. This demonstrates the model’s potential for efficient data labeling tasks.
Another internal testing scenario involved incident response where our on-call engineers use AI to retrieve relevant knowledge to respond to live incidents across logs, ticketing systems, calls, messages, and other data sources. Microsoft-Decision-1 performed better and faster than an LLM in this task.
The model has also been used for scientific discovery by implementing an adaptive replanning feature where an agent evaluates a previous experiment, revises its approach based on rubric grades, and repeats until it completes its objectives.
Microsoft-Decision-1 scored as 46 times more consistent than the LLM-based score at three times the speed and resulted in nearly four times the speed on adaptive replanning. This could significantly impact outcomes for long-running scientific experiments.
The potential applications of Microsoft-Decision-1 are vast, including agent controls, model routing, skill-based decisions, data labeling, AI judging, intent analysis, incident response routing, data validation, recommendations, search relevance, content classification and filtering, code scanning, safety and security screening, computer and UI use, robotics, and scientific discovery.
Developers can get started with Microsoft-Decision-1 today in Microsoft Foundry or through OpenRouter. Input tokens cost $0.042 USD per million tokens, while output tokens are free.
The future of decision models holds great promise for guiding and controlling agents in complex environments. With the introduction of Microsoft-Decision-1, developers now have a powerful tool to tackle fast, reliable decisions.