Microsoft Introduces Decision-1, a Model That Scores Options Instead of Generating Free-Form Text

What Microsoft has announced—and what can be confirmed

Microsoft presents Decision-1 as a decision-scoring model. Rather than necessarily responding with open-ended text, it takes a situation, a question, and a fixed set of possible answers, then returns probabilities associated with those options. That is the defining feature in Microsoft’s model catalog entry.

The difference may sound small, but it changes the kind of task the system is designed for. A generative model typically produces a natural-language response; a scoring model compares alternatives that have already been defined and helps identify which is more likely or suitable within that set. This does not mean the model makes decisions on its own, or that its probabilities guarantee a correct result.

The information available does not confirm all the details usually associated with a product launch: access terms, licensing, usage restrictions, regional or configuration-specific availability, or reproducible evaluation results. It is also not enough to verify the model’s architecture or base model. It is therefore important to distinguish the functional description in the catalog from performance claims that require additional documentation.

Scoring options is not the same as writing a response

In an open-ended interaction, a user might ask a model to explain how to classify a request and receive a detailed justification. That response may be useful to a person, but an automated application would still need to interpret the text, extract a label, and handle possible variations in format. When the options are predefined, a probabilistic output can feed more directly into a follow-up rule: choose a category, send a case for review, or select a tool.

The word “probability” also calls for care. A value associated with an option is a quantitative signal, not a certification that the result will be correct. To use it safely, an organization would need to know how it was calibrated, what data it was evaluated on, and whether that calibration holds in the intended setting. The catalog describes the probabilities as calibrated, but the information provided does not explain the procedure or offer independent results with which to assess it.

The set of answers also limits the scope of the decision. If none of the alternatives represents the case well, the system may be forced to score unsuitable options. A robust workflow should account for a “none of these” outcome, an escalation path, or human review when probabilities are close or insufficient. We cannot assume Decision-1 includes those mechanisms: unless the model documentation says otherwise, they are design choices for the system that integrates it.

Where it could fit in an agent workflow

A software agent often combines several operations: it receives a request, classifies the work, decides which tool to use, and checks whether the result meets certain criteria. A model that scores alternatives may be useful at points where the system already knows the valid options. For example, it could help choose between categories in a support queue or among available tools, provided the application defines the alternatives, validates the output, and handles uncertain cases.

These are possible design patterns, not demonstrated Decision-1 results in the available sources. The catalog describes the scoring format, but does not establish that the model reaches any particular accuracy rate in routing, prioritization, verification, or another specific scenario. Its usefulness will depend on the quality of the options, the information provided, and evaluation in the intended environment.

Microsoft’s general documentation on selecting models for agents offers context for thinking about model choice by task. It is not, however, an evaluation of Decision-1 and does not show that it outperforms other models. That distinction matters when making a purchasing or architecture decision: general guidance can help structure the questions, while a product evaluation needs specific, comparable data.

How to interpret the accuracy and latency figures

The launch proposal refers to a set of 36 tests, an accuracy figure, and latency of 85 milliseconds. The verifiable materials available here do not establish how those tests were built, which models were compared, what counted as a correct answer, or whether results were calculated on the same tasks and under the same conditions. Those figures should therefore not be treated as confirmed general performance.

A useful evaluation would describe the dataset, prompts, number and type of options, scoring procedure, and how results were aggregated. It should also clarify whether the tests represent real applications, whether development and evaluation data were kept separate, and how the model behaves when the correct answer is not among the alternatives. Without that information, an isolated accuracy number cannot predict performance in a particular organization.

Latency also cannot be interpreted without context. To assess 85 milliseconds, readers would need to know, among other things, the hardware, where the measurement was taken, input size, number of options, concurrency, and whether network time was included. A result measured under controlled conditions does not necessarily reflect the time an application will observe in production. The catalog entry consulted does not provide enough detail to confirm that the figure is comparable with those reported for other models.

Access, licensing, and comparisons: questions still open

The documentation provided does not confirm whether Decision-1 is available in Microsoft Foundry, on OpenRouter, or on both, or under what terms. A license, usage limits, specific output options, and deployment requirements have not been verified either. A catalog listing is evidence that Microsoft describes the model there; by itself, it does not settle these operational conditions.

There are also no verified independent evaluations here that compare Decision-1 with alternatives on the same task. General documentation on model selection for agents can help guide an evaluation, but it cannot replace comparable performance tests. Before adding the model to a sensitive workflow, organizations should measure both decision quality and the consequences of errors: a wrong label that delays a service request may have different effects from a recommendation that a person reviews.

To follow related developments, the news index provides access to editorial coverage, while the comparison and discovery sections can help organize alternatives and explore tools. None of these replaces reviewing product documentation or running tests of your own. For now, the prudent conclusion is specific: Decision-1 is presented as an option for scoring fixed answers; its performance, access terms, and advantage over other solutions require more complete documentation and testing.

Keep exploring

Corrections and transparency

If you spot incorrect or outdated information, send us a correction with the page and source we should review.