Microsoft launches new AI model that quickly chooses between answer options

Microsoft has launched a new AI model, Microsoft-Decision-1, built to make fast decisions. It can already be used in Microsoft's developer service Foundry, and it will soon be available on OpenRouter as well.

The model doesn't write long texts like a typical chatbot. Instead, it receives a fixed list of options and responds with a probability for each one. That could be yes/no, multiple choice or a grade. The answer comes in a fixed format, so a program can act on it immediately.

Microsoft says the model can be used for tasks such as sorting tasks, categorizing and prioritizing content, steering workflows, controlling AI agents, scoring other AI responses and running safety checks on content.

Faster than large language models

The model is built by further training Qwen3.5-9B so that it gives answers in a single step. Microsoft plans to later build it on other models, including ones from Microsoft AI and OpenAI.

According to Microsoft's own tests, the model had the highest accuracy in a benchmark with 36 test sets and nearly 150,000 questions it had not been trained on. It was 4.5 times faster than the runner-up, Quyet-1.0-Large, and 35 times faster than GPT-6 Sol (at typical response time).

Microsoft also tested whether the model is stable. When the same request was posed in eight slightly different ways, the answer changed in 1.3 percent of cases on average. When the descriptions of the options were reworded, or the order was reversed or shuffled, the answer did not change at all. Safety tests included 5,250 requests involving, among other things, harmful content, attempts to trick the model and hidden instructions.

Tested internally at Microsoft

Xbox Research used the model to label more than 10,000 pieces of feedback. The quality was on par with GPT-6 Sol, but it was more than 14 times faster and 200 times cheaper. The Copilot team judged it to be on par with GPT-5.6 Luna and 100 times faster.

In a test involving scientific replanning, the assessments were 46 times more consistent than with a method based on a general-purpose language model, and the replanning ran almost four times faster.

Pricing

Input costs from $0.042 per million tokens (tokens are the small chunks of text AI models work in). Output tokens are free.

Microsoft sees the model as a cheap control layer for AI systems. The probabilities can determine whether a program should act, wait, try again, escalate the case or hand the task to a model, a tool or a human.