Microsoft-Decision-1 : 35 times faster than GPT-6 Sol for decision-making

Microsoft-Decision-1 ranks predefined options at $0.042 per million tokens. A targeted alternative for sorting agent workflows.

Microsoft has just introduced Microsoft-Decision-1, an artificial intelligence model designed to make rapid decisions from predefined options. Unlike general-purpose language models capable of writing, conversing, or developing reasoning, this one focuses on precise tasks: classifying information, assigning scores, prioritizing requests, or determining an AI agent's next action.

Developed from Qwen3.5-9B, the model does not seek to produce an elaborate response. Faced with several possibilities, it assigns a probability to each, allowing an application to choose an action or request human verification when the result remains uncertain. This approach specifically targets systems that chain together numerous decisions and for which every additional delay ultimately counts.

Microsoft claims particularly high results. Across 36 tests comprising nearly 150,000 questions, Decision-1 reportedly achieved the highest accuracy among the evaluated models. The company also announces an execution speed 35 times faster than GPT-6 Sol, measured on median latency. These performance metrics stem from evaluations conducted by Microsoft and specifically concern structured decision tasks, without constituting a general comparison of the models' capabilities.

The group is already experimenting with this technology internally. Xbox Research notably used it to categorize over 10,000 player feedback entries, with results deemed comparable to those of GPT-6 Sol, but with processing reported to be 14 times faster and 200 times less expensive. Other trials involve Copilot quality control, IT incident management, and scientific research.

Microsoft-Decision-1 is available in public preview in Microsoft Foundry, as well as via OpenRouter. Its announced price is $0.042 per million input tokens, with no charge for output tokens. Microsoft also plans to evolve its technical foundation toward its own MAI models and those of OpenAI.

With Decision-1, Microsoft advocates an approach in which specialized models handle certain repetitive operations, while general-purpose models remain deployed for tasks requiring more generation or reasoning. A division of labor that could become important as AI agents multiply across professional applications.