Microsoft-Decision-1: A $0.042 Decision Model on Foundry

Microsoft-Decision-1 is a new kind of small model from Microsoft, built to answer one question quickly: which option is correct, and how confident should software be in that answer? Announced on October 9, 2026 on Microsoft's Command Line blog, the decision model is available in Microsoft Foundry and is priced at a level that makes it practical to call thousands of times inside a single agent workflow.

• What it is: a decision-scoring model that returns a calibrated probability for each option in one structured API call.
• Base model: Qwen3.5-9B, post-trained by Microsoft for single-pass decision scoring.
• Price: $0.042 per million input tokens; output tokens are free.
• Speed claim: Microsoft reports P50 latency about 35x faster than GPT-6 Sol.

What Is Microsoft-Decision-1 Built to Do?

Most language models write text. Microsoft-Decision-1 scores choices. Give it a fixed set of options, such as yes or no, a multiple-choice list or a rating scale, and it returns a probability for each. Microsoft positions it for routing, classification, prioritization, verification and workflow control, plus rubric-based grading of AI responses and agent actions.

The post, written by Achint Srivastava, VP of Software Engineering in the Office of the CTO, says the model was made by post-training Alibaba's Qwen3.5-9B. Microsoft adds that it will "soon rebase" the model on others, including its own MAI models and OpenAI models. The post does not describe Microsoft-Decision-1 as open-weight, so treat it as a hosted service.

Calibration is the design goal. Microsoft says the probability is a first-class output, with the aim that a 90% prediction is right about nine times in ten on representative cases.

How Fast and Accurate Is the Decision Model?

These figures are Microsoft's own, so read them as vendor claims. The company reports the highest accuracy in a comparison across 36 benchmarks and nearly 150,000 questions held blind from training; The Register puts that accuracy at 83.5%. On speed, Microsoft says it is 2.5x quicker than the runner-up, H2O-Lightning-4B v1.1.

Robustness is the more interesting number. Across eight types of prompt perturbation, the model changed its decision on 1.3% of cases on average, with zero flips when options were paraphrased, reversed or shuffled. Microsoft also tested 5,250 requests across 11 safety benchmarks.

Inside Microsoft, Xbox Research used it to label more than 10,000 pieces of player feedback at quality the team calls competitive with GPT-6 Sol, while running over 14x faster at roughly 200x lower cost. The Copilot team reports it was about 100x faster for quality control of responses.

Why Do Agents Need Decision Models?

An AI agent makes many small choices: which tool to call, whether an answer passes review, which queue a ticket belongs in. Microsoft's own example makes the point: 100 milliseconds added to each of 20 sequential decisions adds two seconds to a workflow. A model that returns a calibrated score in one pass removes much of that delay and gives developers a number they can set thresholds on.

That is why this category is filling up quickly. We recently covered Liquid AI's d1 open decision models and Jev's calibrated decision model, and Microsoft's benchmark tables include both.

Where Can You Use It?

Microsoft-Decision-1 is live in Microsoft Foundry, with documentation on Microsoft Learn. Microsoft's post also lists OpenRouter; The Register reports that listing is arriving soon, so check availability there before planning around it.

For teams building agents, the takeaway is practical: decision steps no longer need a full-size model. More in our AI coverage.

Sources: Microsoft Command Line — Microsoft-Decision-1 on Foundry — October 9, 2026; The Register — Microsoft's decision model built on an open-weight base — October 10, 2026.