Overview
Released by Microsoft on October 9, 2026, Microsoft-Decision-1 is a decision model that does not chat and does not write prose. It turns an input into a judgment your code can use directly. Input can be text or images, and output comes back as typed answers such as probabilities, options or scores, removing the step of parsing a natural-language answer. Post-trained from Qwen3.5-9B, it runs on Microsoft Foundry and OpenRouter, and Microsoft positions it for incident response, quality control and scientific discovery, where a fast judgment matters more than a fluent paragraph. Thinking of it as a scoring endpoint that reasons is more accurate than thinking of it as a chat model.
Key Features
- Typed output: Returns probabilities, options or scores directly, with no regular expressions needed to extract them
- Low latency: Microsoft puts a single decision at roughly 150 ms, about 10x faster than generating with a general model and parsing
- Stable under perturbation: Microsoft reports about a 1.3% decision flip rate when inputs are perturbed
- Two ways in: Available on both Microsoft Foundry and OpenRouter, so it can replace the judgment step in an existing pipeline
- Input-only billing: Priced at $0.042 per million input tokens with output tokens free and a 32K context
Use Cases
- Batch pipelines that turn text into scores or classifications
- Fast triage in incident response
- Content moderation and quality scoring
- Replacing hard-coded thresholds and rule engines
- Pre-screening before human review
Pros
- Structured output: no parsing or validation step needed
- Fast enough for real-time paths
- Two ready integrations: Foundry and OpenRouter
- Predictable cost: billed on input tokens only
- Stable judgments: low flip rate on perturbed inputs
Pricing
Billed by usage through Microsoft Foundry and OpenRouter at $0.042 per million input tokens, with output tokens free and a 32K context length.
Summary
Microsoft-Decision-1 splits judgment out into its own endpoint: you pass in content and get back a score or option that goes straight into a database, rather than text that needs a second pass to parse. If your pipeline today asks a general model for an answer and then regex-matches it, or hard-codes a set of thresholds, this is worth a comparison run. For conversation, long-form generation or open-ended reasoning, a general-purpose model remains the better fit.
Version History
- Microsoft-Decision-1 lands on Foundry and OpenRouter (2026-10-10): Microsoft released Microsoft-Decision-1, a decision model post-trained from Qwen3.5-9B and built for structured decisions such as probabilities, options and scores, now available on Microsoft Foundry and OpenRouter. Microsoft says it records the highest accuracy across about 36 blind benchmarks and runs 4.5x faster than the runner-up, at $0.042 per million input tokens with free output and a 32K context.