Microsoft-Decision-1: A Model That Only Makes Decisions

Microsoft introduced a narrow decision model, Microsoft-Decision-1, based on Qwen3.5-9B, which outputs calibrated probabilities for a fixed set of options in a single pass instead of generating text; according to Microsoft's measurements, the model outperforms LLMs in accuracy on 36 hidden benchmarks and is 35 times faster than GPT-6 Sol.

Compiled by Sergey KostenchukPublished 2026-10-09Updated 2026-10-10

2026-10-09 Research Microsoft

🤖 Microsoft-Decision-1: A Model That Only Makes Decisions

The narrow Microsoft-Decision-1 does not generate text but outputs calibrated probabilities in a single pass: yes/no, selection from a list, rating, evaluation of AI responses. On 36 hidden benchmarks, the model demonstrated the best accuracy among LLMs and decision models, and according to Microsoft's measurements, is 35 times faster than GPT-6 Sol. The price is $0.042 per million input tokens; available in Microsoft Foundry and OpenRouter.

🌍 Decision models are becoming a pipeline primitive: ready probabilities for routing, labeling, and agent control. Xbox Research labeled over 10,000 reviews — at the level of GPT-6 Sol, but 14 times faster and 200 times cheaper.

👤 The task of "choose one of N" in bots and agent pipelines is solved for a fraction of a cent instead of a large LLM, and the calibrated probability indicates when a human is needed.

Source 1: https://commandline.microsoft.com/microsoft-decision-1-model-foundry/

Sources