Microsoft launched Microsoft-Decision-1 on October 9, targeting the small, repeated choices inside AI applications rather than open-ended conversation. Built on Qwen3.5-9B, it scores predefined options instead of writing an answer.
Microsoft lists pricing at $0.042 per million input tokens, with output free. The model is available in public preview through Microsoft Foundry and also through OpenRouter, whose listing shows the same input and output rates.
It supports yes/no decisions, multiple-choice selection, and rubric-based scoring—useful for routing requests, evaluating responses, and choosing an agent’s next step. The application supplies the context and bounded questions; the model returns numerical decisions and probabilities rather than a free-form explanation.
Why it matters for builders
Not every step in an agent workflow needs a large model generating paragraphs. A dedicated decision layer could:
- Route app-support tickets to the right workflow.
- Choose which model, tool, or agent handles a request.
- Score generated output before it reaches users.
- Escalate uncertain choices for human review.
The practical opportunity is a more modular system: generative models create and explain; decision models handle bounded choices. This follows the architectural pattern behind Cloudflare’s Clef decision models, with a focused scoring component inside the wider agent workflow.
Microsoft’s documentation recommends using probabilities to decide when to act or escalate, while validating performance on your own workload. That means defining the options clearly, testing against representative examples, and choosing thresholds that suit the task. This is a workflow component, not a standalone autonomous agent.