Microsoft announced Microsoft-Decision-1 on October 9 (local time), a model that returns probabilities for predefined choices instead of generating text. It is now available to developers through Microsoft Foundry and OpenRouter. Pricing is $0.042 per 1 million input tokens, with output free.

Microsoft fine-tuned the model from Alibaba Group’s open-weight Qwen3.5-9B. Given a situation, question, and fixed answer choices, it returns each choice’s probability in JSON in a single processing step. The probabilities are calibrated to match actual accuracy, allowing an app to continue automatically or send a case for human review based on those values. The model supports yes-or-no questions, multiple choice, ratings, and scoring AI answers or agent actions against criteria. It has a 32,768-token context window and accepts and returns text only. Microsoft says it plans to move to models based on Microsoft AI or OpenAI models in the future.

Intended uses include classifying, routing, prioritizing, and validating requests; deciding whether to allow an AI agent’s next action; evaluating AI outputs; and labeling data. It does not support text generation, open-ended answers, conversation, translation, or summarization. Microsoft advises against using the model alone to make consequential decisions affecting people, such as credit, hiring, or medical decisions.

Microsoft says Decision-1 recorded the highest accuracy and fastest response speed across 36 benchmarks not used for training. It was 2.5 times faster than the second-place H2O-Lightning-4B v1.1 and 35 times faster than GPT-6 Sol. In a comparison of calibration—the match between predicted probabilities and actual accuracy—it ranked third behind Jev and others. Microsoft also reports that its Xbox research team used the model to classify more than 10,000 pieces of feedback about games and other topics. It maintained quality comparable to GPT-6 Sol while running more than 14 times faster at one two-hundredth the cost.

TypeSafe AI announced Jev, a decision-focused model, on September 15, and OpenAI announced a preview of Decisions API on September 29, offering similar functionality through its low-cost Luna model. Jev also costs $0.042 per 1 million input tokens, with free output, matching Decision-1’s price. Microsoft updated its blog after launch to add comparison results for Jev’s accuracy and calibration.

Microsoft announced Microsoft-Decision-1 on October 9 (local time), a model that returns probabilities for predefined choices instead of generating text. It is now available to developers through Microsoft Foundry and OpenRouter. Pricing is $0.042 per 1 million input tokens, with output free.

Microsoft fine-tuned the model from Alibaba Group’s open-weight Qwen3.5-9B. Given a situation, question, and fixed answer choices, it returns each choice’s probability in JSON in a single processing step. The probabilities are calibrated to match actual accuracy, allowing an app to continue automatically or send a case for human review based on those values. The model supports yes-or-no questions, multiple choice, ratings, and scoring AI answers or agent actions against criteria. It has a 32,768-token context window and accepts and returns text only. Microsoft says it plans to move to models based on Microsoft AI or OpenAI models in the future.

Intended uses include classifying, routing, prioritizing, and validating requests; deciding whether to allow an AI agent’s next action; evaluating AI outputs; and labeling data. It does not support text generation, open-ended answers, conversation, translation, or summarization. Microsoft advises against using the model alone to make consequential decisions affecting people, such as credit, hiring, or medical decisions.

Microsoft says Decision-1 recorded the highest accuracy and fastest response speed across 36 benchmarks not used for training. It was 2.5 times faster than the second-place H2O-Lightning-4B v1.1 and 35 times faster than GPT-6 Sol. In a comparison of calibration—the match between predicted probabilities and actual accuracy—it ranked third behind Jev and others. Microsoft also reports that its Xbox research team used the model to classify more than 10,000 pieces of feedback about games and other topics. It maintained quality comparable to GPT-6 Sol while running more than 14 times faster at one two-hundredth the cost.

TypeSafe AI announced Jev, a decision-focused model, on September 15, and OpenAI announced a preview of Decisions API on September 29, offering similar functionality through its low-cost Luna model. Jev also costs $0.042 per 1 million input tokens, with free output, matching Decision-1’s price. Microsoft updated its blog after launch to add comparison results for Jev’s accuracy and calibration.