Microsoft on Friday released Microsoft-Decision-1, an artificial intelligence model that doesn't write anything. It reads a request, looks at a fixed list of possible answers and picks one, along with how confident it is.
Satya Nadella, Microsoft's chief executive, announced it on X. He said it beat both chatbots and rival "decision models" on speed and quality, and that Microsoft was already testing it "for everything from incident response and quality control to scientific discovery".
Introducing Microsoft-Decision-1, our new model for fast decision-making. It delivers top performance on structured decision tasks, outperforming both LLMs and other decision models in latency and quality. We’re already testing it across Microsoft for everything from incident response and quality control to scientific discovery.
Chatbots like ChatGPT write their answers one word at a time, which is slow and expensive when software needs only a yes or no. A decision model answers in a single step. The answer might be which team should handle a support ticket, whether a message is spam, or whether an AI agent's next action is safe to carry out. Software can act on the result right away. Microsoft says that matters most for agents that make dozens of small choices in a row: adding a tenth of a second to each of 20 steps adds two seconds to the whole job.
The category is only weeks old. A San Francisco start-up, TypeSafe AI, started it on Sept. 15 with a model called Jev. Within a month, more than 90 rivals, most built on free, openly released models, had appeared on JevBench, an independent leaderboard that ranks decision models. OpenAI opened its own version to all developers on Tuesday, running on its cheapest model, GPT-6 Luna. "Pretty interesting how fast everyone deployed decision models," Lisan al Gaib, an AI commentator on X, wrote on Friday.
Microsoft's numbers are strong but come from its own tests. It says the model had the highest accuracy across 36 benchmarks with nearly 150,000 questions it had not trained on. It says the model answered 4.5 times faster than the runner-up, an open model called Quyet-1.0-Large, and 35 times faster than OpenAI's mid-priced GPT-6 Sol. Microsoft did not score Jev, the model that defined the category, on accuracy. It said it compared only against public models. No outside group had published results by Friday evening.
Microsoft did not build the model from scratch. It took Qwen3.5-9B, a small model released for free by the Chinese tech company Alibaba, and trained it further for the job. Many smaller rivals did the same. "The distillation cycle continues," the Wharton professor Ethan Mollick wrote. Microsoft said it would soon rebuild the model on its own technology and on OpenAI's.
The price is the real challenge to OpenAI, Microsoft's closest partner. Microsoft charges $0.042 for every million tokens, the chunks of text AI companies bill by, and nothing for the answer. That matches TypeSafe's price. OpenAI's version costs $0.10 per million.
Inside Microsoft, the Xbox research team used the model to sort more than 10,000 player reviews and social media posts into themes. The company says it was about as accurate as GPT-6 Sol, more than 14 times faster and 200 times cheaper.
The model is available now on Microsoft Foundry, the company's service for developers, and on OpenRouter, a site that offers many AI models in one place. Some developers on Hacker News said on Friday that they could not yet find instructions for calling it.