Microsoft announced Microsoft-Decision-1 in public preview in Microsoft Foundry on October 9, 2026. The model is designed to score defined choices and return structured numerical results—not to write open-ended answers. Microsoft says it post-trained the model from Qwen3.5-9B for single-pass decision scoring. Microsoft’s announcement
Three question types, three kinds of decision
Microsoft’s API documentation defines three question types. A request can provide state as text or JSON, and the response contains numerical decisions without a written rationale. Microsoft’s Microsoft-Decision-1 documentation
| Question type | Task | Returned result |
noul | Evaluate a yes-or-no condition | A probability for the condition |
choice | Select among fixed options | A selected option and probabilities for the options |
score | Evaluate an ordered scale | A probability-weighted score |
That output contract makes the model a fit for bounded tasks such as classification, routing, ranking, evaluation, and workflow control. It does not supply the written explanation behind a choice. If an application needs both a structured decision and a natural-language rationale, Microsoft says developers can combine a decision model with a generative model or another component.
What a Foundry request needs
A developer first deploys the model in Microsoft Foundry, then sends a request to /providers/microsoft/v1/systemone. The request includes the deployment name, a state, and one or more typed questions. Microsoft documents Microsoft Entra ID or an API key for authentication. Microsoft’s API and deployment guide
Microsoft lists GlobalStandard and DataZoneStandard deployment types in selected regions. The documentation cautions that a deployment type does not guarantee lower latency; it recommends testing latency, decision quality, and confidence calibration on workloads representative of the intended application.
Scores still need oversight
Microsoft warns that scores can change with the wording of a question or the order of its options. Calibration is strongest on familiar task types, and the model may rely on outdated knowledge. Its safety filters can miss subtle harmful content or flag benign content. Microsoft’s usage guidance
Microsoft says the model should not be the sole basis for consequential decisions about individuals. For areas such as credit, employment, housing, health care, and legal decisions, the company recommends meaningful human review.
Fahd Mirza’s individual prompt demonstrations include security, routing, classification, and prompt-injection scenarios. Mirza reports mixed results on keyword-based routing; a handful of examples does not establish how the model will perform across other prompts or workflows.
How to read Microsoft’s speed claim
Microsoft’s reported comparison says Microsoft-Decision-1 was 35 times faster than GPT-6 Sol. The figure is a company-reported benchmark claim relayed by The Next Web. Microsoft’s documentation points developers back to testing their own representative workloads for latency, decision quality, and confidence calibration.
