Microsoft builds Decision-1 on Alibaba's Qwen and plans to move it to OpenAI and MAI models

The model scores a fixed list of choices instead of writing answers, and Microsoft puts it 35 times quicker than GPT-6 Sol.

By Think Facility · October 11, 2026 · Microsoft Open models

On October 9, 2026, Microsoft released Microsoft-Decision-1, a small model that picks an answer from a list you give it and says how sure it is. It's in public preview in Microsoft Foundry.

Microsoft didn't train it from scratch. Its launch post says the team post-trained Qwen3.5-9B, an open model from the Qwen team at Alibaba Cloud, and plans to move it onto Microsoft's own MAI models and OpenAI's.

we post trained Qwen3.5-9B for fast, single-pass decision scoring and will soon rebase it on other models, including Microsoft AI (MAI) and OpenAI

From Microsoft-Decision-1: Our model for fast decision-making

What a decision model does

It doesn't write text. You hand it a request and a closed set of options (yes or no, a team to route to, a score on a scale), and it returns a calibrated probability for each. An app can then act, wait, or pass the case to a person.

A blue Microsoft Research banner on a lamppost, seen from below, between two pale stone office buildings with long rows of windows — Microsoft Research's building 99 in Redmond, Washington. Photo: Robert Scoble, CC BY 2.0, via Wikimedia Commons

The Foundry announcement lists the jobs it's meant for, from agent controls and model routing to AI judging and content classification. The pitch is cost. Most steps inside an agent are small choices, and Microsoft would rather you didn't pay a big generative model to make them.

Model: Microsoft-Decision-1
Built on: Qwen3.5-9B
Released: October 9, 2026, in public preview
Input tokens, per 1M: $0.042
Comparison: 36 benchmarks, nearly 150,000 questions

How Microsoft says it scores

Achint Srivastava, a vice president in Microsoft's Office of the CTO, wrote the launch post. He puts the model first on accuracy across the company's own comparison.

Microsoft-Decision-1 achieved the highest accuracy in our 36-benchmark comparison, spanning nearly 150,000 questions across benchmarks kept blind from training.

From Microsoft-Decision-1: Our model for fast decision-making

On speed, the post puts it 2.5 times quicker than the runner-up, H2O-Lightning-4B v1.1, and 35 times quicker than GPT-6 Sol. Microsoft also checked how steady its answers are: changing a request in eight ways flipped the decision on 1.3% of tries, on average.

These are Microsoft's numbers, from a comparison Microsoft ran. The post also carries an editor's note saying benchmarks for Jev, another decision model, were added after it first went up.

Who has used it

Inside Microsoft, Xbox Research sorted more than 10,000 pieces of open-ended feedback with it, and the Foundry post says quality was "competitive with GPT-5" at 80 to 100 times the speed. Other teams tested it for grading Copilot answers.

Input costs $0.042 per million tokens in both the US and EU data zones. The Foundry pricing table puts N/A in the output column, and since the model answers with a handful of probabilities, input is most of what you'd send it anyway.