Microsoft-Decision-1 Operational

Microsoft-Decision-1 is Microsoft's next-generation fast decision-making AI model, designed for structured decision tasks. It selects the optimal solution from predefined options and assigns probability scores to each alternative for downstream applications to decide whether to execute, retry, or transfer to a human. Microsoft-Decision-1 ranked first in accuracy across all 36 benchmarks, is 35 times faster than GPT-6 Sol, and costs only $0.042 per million input tokens, with free output. The model is built on Qwen3.5-9B and is available on Microsoft Foundry.

Optimal solution selection: Selects the best answer from predefined options and outputs probability scores for each alternative.
Routing and distribution: Intelligently routes requests or tasks to the appropriate downstream modules or services.
Content classification: Quickly categorizes text and other content; for example, the Xbox team uses it to classify tens of thousands of game feedback items.
Task prioritization: Determines the processing order for multiple pending tasks.
Result verification: Evaluates the quality of AI-generated results and decides whether to approve or retry them.
Workflow control: Acts as an orchestration component in Agent workflows, determining whether to continue execution, retry, escalate, or transfer to a human.

Base architecture: Microsoft-Decision-1 is built on the open-source Qwen3.5-9B model. Its moderate parameter size balances inference speed and deployment cost. Microsoft plans to migrate it to other base models in the future, including its proprietary MAI: series and OpenAI models.
Specialized post-training: The model underwent specialized post-training centered on “single-pass decision scoring,” using probabilistic scoring for predefined options to transform a large language model into a structured decision engine.
Decision mechanism: The model completes scoring and selection for each decision request in a single pass, substantially reducing latency compared with large language models that use multi-step reasoning. According to Microsoft, the cumulative latency issue in which 20 consecutive decisions each add 100 milliseconds is significantly compressed by this architecture.
Robustness design: The model is specifically optimized for decision stability. Across eight perturbed versions of the same request, the average decision-flip rate is only 1.3%, and the flip rate is zero when option descriptions are rewritten or their order is reversed, ensuring that input changes do not affect decision consistency.
Safety alignment: The model underwent safety testing covering 11 benchmarks and 5,250 requests, including harmful-content detection, jailbreak attacks, and prompt injection. It maintains high utility on normal tasks while refusing harmful requests.

Connect to the platform: Developers can access the model through Microsoft's AI development platform, Microsoft Foundry.
Define the decision scenario: Convert the task into a structured decision problem consisting of “predefined options + optimal-solution selection,” such as routing, content classification, or priority ranking.
Submit a request: Submit the content requiring a decision to the model in a structured format.
Obtain probability scores: The model returns the optimal option and probability scores for each alternative instead of open-ended text.
Drive downstream actions: Use the scores to determine whether to continue execution, retry, escalate, or transfer to human review.
Embed in workflows: Seamlessly embed the model into existing applications, AI Agents, and workflow systems as a foundational orchestration component.

Extremely fast: P50 latency is 35 times faster than GPT-6 Sol and 4.5 times faster than Quyet-1.0-Large, making it the fastest decision model in testing.
Leading accuracy: Ranked first in accuracy across 36 benchmarks covering nearly 150,000 questions.
Very low cost: Input costs only $0.042 per million tokens, with free output. Internal testing found that classification task costs were approximately 200 times lower than GPT-6 Sol.
Scored decisions: Outputs probability scores for each alternative, making it easier for downstream systems to decide whether to execute, retry, or transfer to a human. This is more controllable than “black-box” generation.
Strong robustness: The average decision-flip rate is only 1.3% across eight perturbations, and the flip rate is zero when options are rewritten or reordered.
Safe and reliable: Passed safety testing covering 11 benchmarks and 5,250 requests, refusing harmful requests while maintaining utility for normal use.

Project website: https://commandline.microsoft.com/microsoft-decision-1-model-foundry/

Intelligent routing and distribution: Intelligently routes user requests or task traffic to the most suitable downstream service, model, or Agent.
Content classification and moderation: Quickly classifies large volumes of user feedback and comments and detects harmful content.
Task prioritization: Determines task processing order in scenarios such as customer service tickets and defect handling, enabling automated workflow orchestration.
AI result verification: Acts as an AI response evaluator to determine whether generated results meet requirements and whether to approve, retry, or escalate them.
Exception decisions and escalation: In scenarios such as incident response and quality control, uses probability scores to determine whether to transfer cases to human review, creating a safe human-AI collaboration fallback mechanism.