Microsoft's Decision-1 Scores Choices Instead of Writing Them

Not every job inside an AI product needs a model that can write an essay. A lot of the work is smaller and more mechanical: should this request go to the big model or the cheap one, did the agent's last step actually succeed, is this ticket urgent or not. Microsoft's new model is built for that unglamorous middle layer. Reported on October 9, Microsoft-Decision-1 does not produce prose. It takes a fixed set of options and returns a calibrated probability for each one.

The model is post-trained on Alibaba's Qwen3.5-9B, according to coverage from MarkTechPost. It handles yes or no questions, multiple choice, ratings, and rubric grading of an AI's own responses or an agent's actions. Because the output is a probability rather than free text, developers can set thresholds and wire the result straight into logic. One common pattern: auto-approve anything the model scores above 90 percent confidence, and send the rest to a person. It carries a context window of 32,768 tokens.

The pitch is price

Decision-1 is cheap by design. Microsoft lists input at 0.042 dollars per million tokens, with output free. For comparison, OpenAI's Luna decisions endpoint charges 0.10 dollars per million input tokens. On a workload of a million short tickets, one analysis put the running cost at roughly 12.60 dollars. That is the sort of figure that makes it practical to run a classifier on every request rather than sampling.

It lands in a corner of the market that is suddenly busy. Liquid AI recently released its own small models that return an answer in a single pass, covered here when Liquid AI opened its d1 decision models, and OpenAI has a Decisions API of its own. The shared idea across all of them is that fast, confident scoring is worth separating from slow, expensive generation.

Read the fine print

The performance figures come from Microsoft, not an outside lab. The company reports accuracy of 83.5 percent and a median latency of 85 milliseconds across its own 36-benchmark comparison, numbers that have not been independently checked. The model is available through Microsoft Foundry and needs an Azure account; reports on whether it is listed on OpenRouter yet are inconsistent. And unlike the open-weight options it competes with, Decision-1's weights stay with Microsoft, so there is no self-hosting it. Microsoft has said it plans to rebase the model on other foundations later, including its own MAI models and OpenAI's.

For teams already running a fine-tuned, BERT-class classifier, Decision-1 is not an obvious upgrade so much as a convenient managed alternative. For everyone bolting decision logic onto agents without wanting to train anything, a model that scores choices for pennies is a tidy fit.