Microsoft has released Microsoft-Decision-1, a hosted model built to score choices rather than generate text. Where a normal chat model writes an open-ended reply, Decision-1 takes a description of a situation plus a fixed list of options and returns a calibrated probability for each one. That covers yes/no questions, multiple-choice picks, numeric ratings and rubric-based grading of an AI agent's own actions or responses.
The model is a post-trained version of Alibaba's open-weight Qwen3.5-9B, released on 9 October and available now through Microsoft Foundry and OpenRouter. It works with a 32,768-token context window and always returns JSON, which suits it to sitting inside an automated pipeline rather than a chat interface. Microsoft is not releasing the weights; this is an API-only product.
Pricing is unusually simple: $0.042 per million input tokens, with output tokens free, reflecting how short Decision-1's replies are by design. Microsoft reports a median response time of around 85 milliseconds and says the model scored the highest average accuracy across 36 internal benchmarks covering close to 150,000 questions. Those figures are Microsoft's own and have not been independently checked.
The use case is agent infrastructure: something has to decide which of several candidate actions an AI agent should take next, or grade whether its last answer met a rubric, and doing that with a full-size language model is slow and expensive. Microsoft says it plans to rebase Decision-1 on other models, including its own MAI line and OpenAI's, which suggests this is meant as a durable piece of plumbing rather than a one-off release.
Source: MarkTechPost
