Microsoft Decision-1: 35x Faster Probabilistic Classification via Qwen3.5-9B on Foundry

Microsoft has launched Decision-1, a specialised artificial intelligence model designed to handle structured decision-making tasks without generating free-form natural language. According to InfoWorld, the model returns calibrated probabilities for predefined answer sets, allowing software applications to consume outputs directly. This architecture separates classification logic from the generative capabilities of large language models, targeting use cases such as request categorisation and incident prioritisation.

Root cause

The underlying mechanism differs from general-purpose large language models by eliminating the generation of explanatory text. Microsoft stated in a blog post that Decision-1 takes a fixed set of answer options and returns a probability for each in a single pass. This design reduces computational overhead associated with token generation for output text. The model was constructed by post-training the Qwen3.5-9B base model, with plans to rebase it on Microsoft AI and OpenAI architectures in future iterations.

Attack path

This release does not involve a security vulnerability or attack vector. Instead, it represents an architectural shift in how AI agents process logic. Developers can use the model to classify inputs, rank work items, verify responses, or determine the next action in an automated workflow. Microsoft noted that the model can evaluate multiple decisions in parallel within a single request, which may reduce the surface area for latency-based bottlenecks in agentic systems.

Affected versions

The model is currently available through Microsoft Foundry and OpenRouter. Microsoft stated that Decision-1 achieved the highest accuracy in a comparison of 36 benchmarks covering nearly 150,000 questions that were kept blind from training data. The company reported performance metrics indicating the model is approximately 2.5 times faster than the runner-up and 35 times faster than GPT-6 Sol, though these claims have not been independently verified.

Mitigation

• Monitor integration points where Decision-1 outputs drive critical automated actions to ensure probability thresholds are set correctly.
• Validate that the fixed answer sets provided to the model cover all possible edge cases for your specific workflow.
• Review logging mechanisms to capture the input tokens and resulting probabilities for audit and debugging purposes.

What to do and how to stay safe: Microsoft

• Audit current large language model usage to identify tasks that rely on simple classification or binary decisions.
• Test Decision-1 in a non-production environment to compare latency and cost against existing general-purpose models.
• Define strict input validation rules to prevent prompt injection or unexpected data formats from reaching the model.
• Establish monitoring alerts for abnormal probability distributions that might indicate model drift or data quality issues.

Step-by-step guide: Password Policies Examples That Reduce Password Fatigue

General security guidance from the Vector Bulletin newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

What is the base model for Microsoft Decision-1?

Microsoft built Decision-1 by post-training the Qwen3.5-9B model, with plans to rebase it on other architectures later.

How does Decision-1 differ from a standard large language model?

It returns probabilities for predefined answer options in a single pass rather than generating free-form natural language responses.

What is the cost structure for using Decision-1?

Microsoft lists the price at 4.2 cents per million input tokens, with no charge for output tokens, according to the blog post.