Microsoft launches Decision-1, a non-generative decision-making model designed to score options and respond in only 85 milliseconds

Microsoft launched the Microsoft-Decision-1 decision-making model in Microsoft Foundry on October 9, which specifically handles classification, option judgment and scoring. This model does not use the text answer method of general generative AI. Instead, it calculates the probability score of each option based on the input situation and pre-provided options, allowing the application to decide the next step. The third-party AI model platform OpenRouter was also launched simultaneously, with the model codename "microsoft/microsoft-decision-1".

The purpose of Decision-1 is different from the general chat model. Developers can provide a piece of user information, ask the model to determine their intent, and then transfer the results to customer service, technical support, or other processing processes; it can also be used for data classification, event triage, content security marking, and scale scoring. The model supports true-false questions, multiple-choice questions, and an abstain option when a judgment cannot be made from existing information. Since the output is limited to decision scores, there is no need to generate additional text explanations, making it suitable for integration into automated systems that require rapid judgment.

Microsoft designed Decision-1 as a text-only model with a maximum context length of 32,768 tokens and does not accept images, audio, or videos. The model weights are not disclosed, and developers must use it through the managed API. The billing is $0.042 per million input tokens, and there is no additional charge for output. Microsoft Foundry provides a Decision type API, and deployment options include GlobalStandard and, in some regions, DataZoneStandard.

In terms of model source, Decision-1 is not completely trained from scratch by Microsoft's own model, but is based on Alibaba's open weight Qwen3.5-9B and then post-trained by Microsoft. Foreign media "The Register" pointed out that Microsoft has not yet explained the reasons for its initial selection of Qwen. Microsoft said that subsequent plans will use its own MAI or OpenAI model as the underlying architecture. Although Decision-1 was developed using an open weight model, Microsoft did not make the model weights available upon completion.

Microsoft CEO Satya Nadella also introduced this model on the social platform X, saying that the company has used related technologies internally for incident response, quality control and scientific discovery. The officially announced test covers 36 benchmarks and 147,137 questions not included in training, with an average accuracy of 83.5% and a median response delay of about 85 milliseconds.

Internal Microsoft teams such as Xbox Research, Copilot and Discovery also report that Decision-1 can increase processing speed and reduce costs while maintaining similar quality. However, these performance and cost data come from Microsoft's own testing and have not been verified by an independent third party. The option design, input length and judgment threshold of different tasks may also affect the actual deployment results.

There are currently other players investing in the decision-making model market. After OpenAI launched the Decisions API, there are more than a hundred similar models on the market. Microsoft provides Decision-1 as an independent API this time, allowing developers to handle decision scoring separately from general text generation work. For example, first determine the request type and then decide whether to hand it over to other models or programs for execution.