Microsoft releases Decision-1, a 9B model that scores fixed answer options

Microsoft announced Microsoft-Decision-1 on 9 October 2026. It is a text-only decision-scoring model post-trained from Qwen3.5-9B. You give it evidence, a closed question and a fixed set of answer options. It returns a probability for each option in one pass. It does not write explanations.

Microsoft lists routing, classification, prioritisation, verification and workflow control as uses. The Foundry card adds grading AI responses against criteria, groundedness checks, content-safety flagging with your own thresholds, and an option to abstain when evidence is thin.

Access is through Microsoft Foundry and OpenRouter. The launch post gives a price of $0.042 per million input tokens, with free output. Context is 32K tokens. The Foundry card shows a Preview lifecycle. Neither page states a licence, and neither mentions weights.

The headline numbers are Microsoft's own: top accuracy across 36 blind-held benchmarks and a 1.3% average decision-flip rate under perturbation. The post says it was updated to add JevBench results. Microsoft says the model should not be the sole decision-maker on credit, employment, housing, insurance, education, healthcare or legal rights.

It does what the card says. The card says a careful thing about people, and the benchmarks are the vendor's.

Sources

• https://commandline.microsoft.com/microsoft-decision-1-model-foundry/
• https://ai.azure.com/catalog/models/Microsoft-Decision-1
• https://www.productbuilder.net/models/microsoft-decision-1

Editor's first pass (AI-assisted)

• Real 3.5: Hosted access exists through Foundry and OpenRouter, but the Foundry card lists a Preview lifecycle and weights are not mentioned.
• Useful 4: Routing, classification, grading and agent-action gates are real tasks, and output is cheap per the posted price.
• Notable 4: A large vendor shipping a model explicitly sold as a decision model puts the category on the map.
• Novel 3: The form factor follows other decision-model releases this autumn; the base is an open Qwen model.