Microsoft introduced Microsoft-Decision-1 on October 9, 2026. Built by post-training Qwen3.5-9B, the model targets structured decisions rather than free-form text. The company reports leading results in its own comparison across 36 benchmark sets. These are vendor findings that SIExpo has not independently reproduced.
Where a decision model fits
Microsoft Foundry's announcement identifies the service as a public preview. Listed applications include classification, request routing, and assessment against criteria. The published Foundry price is $0.042 per million input tokens. Those conditions concern the described service, not a promise of identical final pricing through every intermediary.
Our editorial interpretation is that this is a different application building block. A person asking how to solve a complex problem may need an explanatory answer. Software that already knows its permitted next steps may instead need a specific choice. Those tasks should not automatically be judged by the same quality criterion.
In our hypothetical example, a support request must first reach the correct team. A well-written paragraph that cannot serve as a routing decision does not complete that step. Conversely, a correct category is insufficient to compose a useful response to the customer.
How the question is defined
Microsoft Learn's official guide describes an input state containing text or JSON and typed questions. Category choices, binary probabilities, and scores on ordered scales produce different results. The model does not return a freely written explanation. A priority score can be a probability-weighted average, so it should not automatically be portrayed as an absolute measured quantity.
The application needs understandable categories. In our delivery example, we would distinguish a timing enquiry from a damaged-item complaint. Overlapping categories can make even a convincing probability unhelpful. In that situation, the issue starts with the task definition rather than necessarily with model size.
There is more than one access route
Vercel also announced Decision-1 access through AI Gateway on October 9. It lists the AI SDK decision API and compatible decision interfaces. This route does not require a separate Azure deployment; it uses Gateway authentication and the model identifier microsoft/microsoft-decision-1. Calls appear in logs and count toward budgets.
That does not make the integrations interchangeable in every respect. Teams need to consider the provider, authentication, data handling, and cost accounting as part of the chosen route. We are not making paid requests or comparing their actual latency in this article.
Confidence needs a testable meaning
The scikit-learn documentation explains calibration as correspondence between predicted probabilities and observed frequencies. For a well-calibrated binary model, roughly 80% of cases assigned a probability near 0.8 should be positive. Calibration curves compare these quantities across prediction groups. Overall accuracy alone does not demonstrate reliable probabilities.
Its separate prediction-evaluation guide distinguishes accuracy, precision, recall, and other measures. Different classification errors have different meanings. Readers should therefore not treat one overall score as a sufficient description of behavior in every category.
Our example contrasts low-risk routing of an informational enquiry with a consequential change to a customer account. Equal numerical confidence does not automatically imply an equal execution threshold. An organization must define what is allowed after the decision and what remains subject to review. Probability is a signal; authority to act is a separate question.
What a useful deployment would demonstrate
Our editorial assessment would begin with representative cases, including ambiguous and mixed requests. We would then examine category-level errors, end-to-end completion time, and cost per actually finished task. These questions describe an evaluation approach, not a benchmark we have performed.
The news is a concrete expansion of developers' choices: a specialized model for bounded decisions that can become part of a larger workflow. Whether it suits a workload better than an existing solution depends on that workload. The relevant criterion is observable behavior in the task, not the impressive wording of an overall ranking.
