When an agent has to choose its next action or route a request, calling a general-purpose language model for every decision can add latency and cost. Microsoft Decision-1 is designed to score a fixed set of options for these tasks. Microsoft introduced it on October 9, 2026 (U.S. date; October 10 in Korea), and it is now available through Microsoft Foundry and OpenRouter.
Instead of writing a long response like a chatbot, Decision-1 returns calibrated probabilities for a supplied situation and a closed set of choices. It supports yes/no questions, multiple choice, ratings, and rubric-based evaluation of AI responses or agent actions. Its intended uses include routing, classification, prioritization, and verification—workflows where software can act directly on the result.
A single-pass model that scores options
Microsoft says it post-trained Decision-1 from Alibaba’s Qwen3.5-9B. The Foundry model card says training used public datasets and synthetic data created by Microsoft’s team. This does not mean that Microsoft has released Decision-1’s own weights.
Provide the situation to evaluate along with the available answers. In a single pass, the model returns a structured probability for each option. An application can use those scores to choose a tool or send an uncertain case to a person for review.
For example, an agent’s proposed action could be scored as “proceed,” “retry,” “stop,” or “escalate to a person.” The same pattern can route incoming requests to the right model or grade generated output against a defined quality rubric.
The 35× figure is a P50 latency result in Microsoft’s evaluation
In its official announcement, Microsoft says Decision-1 achieved the highest accuracy in its comparison across 36 blind benchmarks and about 150,000 questions. The evaluation covered routing, ranking, long-context tasks, multilingual tasks, out-of-distribution data, reasoning, and safety.
The key speed metric in the announcement is median latency, or P50. In Microsoft’s own comparison, Decision-1’s P50 latency was about 35× shorter than GPT-6 Sol’s and 4.5× shorter than that of Quyet-1.0-Large, the next-fastest comparison model. This describes latency in that evaluation; it does not mean every input or deployment will have 35× the throughput or generation speed.
The announcement does not provide every detail of the comparison environment, such as hardware and region, concurrent request counts, repeat counts, or per-input conditions. Treat the 35× figure as a result reported by Microsoft for its own benchmarks. In a separate robustness evaluation, Microsoft reports that changing input phrasing in eight ways changed the selected result by 1.3% on average, while changing option descriptions or reversing and shuffling their order did not change the selection.
Microsoft’s announcement page includes classification and computer-use demonstrations.
Source: X account @satyanadella
Source: X account @msdev
Built for classification and verification, not open-ended generation
Decision-1 accepts text and supports up to 32,768 tokens of context. It is not designed to write explanations or answer open-ended questions. It is also a poor fit for tasks that need generated prose—such as conversation, summarization, or translation—or knowledge that is absent from the input.
The model card also says not to use Decision-1 as the sole automated decision-maker for consequential decisions involving people, including credit, employment, housing, insurance, education, healthcare, or legal matters. The application remains responsible for setting score thresholds and providing paths to hold uncertain cases or escalate them for human review.
Get started with Foundry or OpenRouter
In the Microsoft Foundry catalog, find Microsoft-Decision-1, deploy it to a project, then call it from the Playground or an API. The Foundry introduction describes it as public preview, while the catalog model card shows GA. Check the catalog for the availability and deployment terms on your own subscription.
The model is also available on OpenRouter under the ID microsoft/microsoft-decision-1. Microsoft’s announcement initially said OpenRouter support was coming soon; the model page is now live, so availability followed after the announcement.
Microsoft’s announcement and OpenRouter list a price of $0.042 per million input tokens, with output tokens free. Foundry pricing can vary by region and deployment, so check the price shown for your subscription before deploying. Both routes list a context limit of 32,768 tokens.
Related reading
For other structured decision-model approaches, read OpenAI Decisions API: Jev Selects Options with GPT-6 Luna, Unsloth Laya Decision Model: Run Locally with 4GB RAM, and CLM-8B: System One’s Jev Decision Model.