Microsoft introduced Microsoft-Decision-1, a specialized artificial intelligence model that evaluates predefined response or action options. According to the company, in tests for making structured decisions, its response latency was 35 times less than that of OpenAI GPT-6 Sol. The model is designed to classify information, validate results, and manage work processes.
Image source: microsoft.com
Microsoft-Decision-1 is not designed to freely generate text like a regular chatbot. Given a question and a set of options, she assigns a probability score to each. The application determines what happens next: for example, it can allow the AI agent to perform a suggested step, require a retry, or delegate the task to a human.
Working with the model is inexpensive: $0.042 for 1 million input tokens, and output data is not charged. This design is especially beneficial for applications that frequently need to classify requests or select next actions. Microsoft-Decision-1 can also be used to route queries to appropriate models, prioritize tasks, and evaluate the quality of responses from other AI systems.
According to Microsoft, the new product showed the greatest accuracy among the compared models across a set of 36 benchmarks, covering almost 150,000 questions. The items from these tests were not used during training. The median response delay was 35 times less than that of GPT-6 Sol, and 2.5 times less than that of H2O-Lightning-4B v1.1, which took second place in this indicator. These results were obtained in tests by Microsoft itself and relate to tasks of choosing from given options.
The company has already tested the model in several practical scenarios. Xbox Research engineers used it to distribute more than 10,000 player reviews into predetermined topics. In terms of classification quality, Microsoft-Decision-1 was comparable to GPT-6 Sol, but was more than 14 times faster at a cost approximately 200 times lower. In tests by the Microsoft Copilot team to assess the quality of answers, it showed a result comparable to GPT-5.6 Luna, with a hundredfold speed advantage.
Microsoft-Decision-1 is built on Alibaba Qwen3.5-9B, which has undergone additional training to evaluate options in a single pass. In the future, Microsoft intends to create versions based on other models, including its own MAI and OpenAI systems. The company also tested the robustness of solutions to changes in the wording and design of queries: in these tests, the model changed the answer on average 1.3% of the time, and there were no changes when rephrasing options or rearranging their order.
To assess security, Microsoft used 11 benchmarks with 5,250 requests, including malicious content, attempts to bypass security restrictions and inject foreign instructions. According to the company, the model rejected malicious requests while maintaining high utility during legitimate tasks. Microsoft-Decision-1 is available now on the Microsoft Foundry platform and through OpenRouter.
If you notice an error, select it with the mouse and press CTRL+ENTER.
