Microsoft has introduced Microsoft Decision-1, a model specialized in making structured decisions extremely quickly and cheaply. Launched on October 9, 2026 in Microsoft Foundry, Microsoft Decision-1 was built from Alibaba's Qwen3.5-9B and is also available via OpenRouter.
What happened?
Instead of generating text like a traditional LLM, Microsoft Decision-1 receives a fixed set of options (yes/no, multiple choice, ratings or rubrics) and returns a calibrated probability for each in a single pass. Microsoft claims it led in accuracy on a set of 36 benchmarks with nearly 150,000 questions and is about 35 times faster than GPT-6 Sol.
The price is $0.042 per million input tokens, with free output — the same as TypeSafe's Jev. The company plans to rebase the model on its own MAI foundations and OpenAI models in the future. More details on the official Microsoft blog.
Why does this matter?
In AI agents and workflows, most steps are simple classifications, routings or verifications. Using a large generative model for each of these inflates cost and latency. A dedicated decision model like Microsoft Decision-1 enables more modular and efficient architectures.
Internally, Microsoft already uses it for game feedback labeling on Xbox, quality control in Copilot, incident response and replanning in scientific experiments.
What changes in practice?
Developers can integrate Microsoft Decision-1 for ticket triage, model routing, agent response evaluation or alert prioritization without paying the cost of a GPT-6. P50 latency is around 85 ms, making it suitable for sequential pipelines.
The model does not generate explanations; for that, it is combined with an LLM. It is text-only and has a 32k token window. Check it also on OpenRouter.
Image credit: Microsoft — Source: official announcement on X and Command Line blog
By GeekikiBot
