Microsoft Decision-1: A Fast, Cheap Model That Only Decides
3 min readMicrosoft has launched Microsoft Decision-1, a small, fast AI model that does not chat or write essays. It picks answers. Give it a fixed set of options and it returns a calibrated probability for each one, and Microsoft says it does this about 35 times faster than OpenAI’s GPT-6 Sol at a fraction of the cost.
What Is a Decision Model?
Large language models are built to generate text and reason through open-ended problems. Decision models are narrower. They return structured outputs that software can act on directly: yes or no, multiple choice, a rating, or a grade against a rubric. That makes them useful for the plumbing of AI systems, such as routing requests, classifying tickets, screening content, or judging whether an agent’s action is safe to run.
The category is crowding fast. The Register counts more than 100 decision models, including TypeSafe’s Jev, Liquid AI’s d1, Cloudflare’s Clef and H2O-Lightning-4B, and OpenAI and Perplexity offer decision APIs of their own.
What Microsoft Decision-1 Offers
Decision-1 is available in Microsoft Foundry and through OpenRouter. Pricing is aggressive: $0.042 per million input tokens, with output tokens free. Microsoft says it was the fastest model in its own testing, about 2.5 times quicker than H2O-Lightning-4B, and that it scored the highest accuracy across 36 benchmarks covering nearly 150,000 questions. It also changed its answer on only 1.3% of deliberately altered inputs on average.
Internal results are striking. Xbox Research ran more than 10,000 pieces of game feedback through it with quality competitive with GPT-6 Sol, while running over 14 times faster and about 200 times cheaper. Achint Srivastava, a VP in Microsoft’s Office of the CTO, called decision models “quickly emerging as an important new category in AI.”
One detail stands out. The first version is post-trained from Alibaba’s open-weight Qwen3.5-9B. Microsoft says it will soon rebase Decision-1 on models from Microsoft AI and OpenAI, but has not said why.
Why It Matters
As companies wire AI agents into real workflows, most of the calls those agents make are small judgments, not essays. Paying frontier-model prices and waiting on frontier-model latency for every one of them adds up fast. A cheap, fast model that only decides could become the traffic cop for agent systems, including the kind of outside controls Microsoft has been arguing for.
The caveat is that decision models still make mistakes, and errors in a routing or safety layer can cascade. Watch for independent benchmarks and for which base model Microsoft lands on next.