Microsoft Releases Decision-1: A Qwen3.5-Based Model for Fast, Calibrated Structured Decisions

Microsoft Releases Microsoft-Decision-1: An AI Model Specialized for "Multiple Choice" Tasks, Post-Trained on Qwen3.5-9B

Microsoft has introduced a new model, Microsoft-Decision-1. It doesn't write articles or chat; it does one thing only: given several options, it quickly determines which one to choose and assigns a probability score to each option. Satya Nadella announced the model on X. It is now live on Microsoft's AI development platform, Microsoft Foundry, and will soon be available on OpenRouter (an API platform aggregating multiple models).

Microsoft calls this type of model a "decision model." Large language models excel at generating text and reasoning through complex problems, but many AI calls in software are actually just making judgments: Which category does this user feedback belong to? Which model should handle this request? Should the agent continue, retry, or escalate to a human? The answers to these questions lie within fixed options, and using large models to answer them is both slow and expensive. Decision models directly output structured results that programs can use immediately.

It was post-trained based on Alibaba's open-source model Qwen3.5-9B and produces results in a single pass. It supports yes/no questions, multiple-choice questions, and scoring, and can also grade AI responses or agent actions according to rubrics. Microsoft states that they plan to retrain it on their own MAI models and OpenAI's models in the future.

The probabilities it outputs are "calibrated": when the model says it is 90% confident, it is correct approximately nine out of ten times. Developers can set thresholds based on this: high confidence triggers automatic execution, while low confidence routes to manual review.

All performance data below comes from Microsoft's internal tests. In comparisons across 36 benchmarks and nearly 150,000 questions, it achieved the highest accuracy. In terms of speed, it is 4.5x faster than the second-place Quyet-1.0-Large and approximately 35x faster than GPT-6 Sol. Microsoft also tested stability: applying eight types of modifications to the same request, such as rephrasing or shuffling option order, resulted in an average change of only 1.3% in judgment outcomes, with zero changes observed when shuffling option order.

Speed matters because agents often need to make dozens of small decisions consecutively during operation. If each decision is delayed by 100 milliseconds, waiting time increases by 2 seconds over 20 steps.

Microsoft is already using it internally. The Xbox research team used it to categorize over 10,000 player feedback entries and reviews by topic, achieving quality comparable to GPT-6 Sol but with 14x+ speed and 200x lower cost. The Copilot team uses it to score response quality, achieving results close to GPT5.6 Luna but with 100x faster speed.

For developers, it is suitable for stages where answers lie within limited options, such as model routing, intent recognition, content moderation, data labeling, acting as a judge for AI outputs, and selecting the next click target when AI controls computers. Replacing large models with this one in these stages can reduce both latency and costs.

Pricing is $0.042 per million input tokens, with free output. It is currently callable via the model catalog in Microsoft Foundry.

↳ Quote @satyanadella (Satya Nadella): Introducing Microsoft-Decision-1, our new model for fast decision-making.

It delivers top performance on structured decision tasks, outperforming both LLMs and other decision models in latency and quality.

We’re already testing it across Microsoft for everything from incident response and quality control to scientific discovery.