What is Microsoft-Decision-1 and why does it matter for software that needs to decide fast?

What makes a decision model different from an LLM

Large language models generate text. They reason through open-ended problems and produce prose that humans read. Decision models do something else: they emit structured outputs that code can consume without parsing. The announcement frames this as a new category in AI. Routing a support ticket, classifying a document, prioritizing a queue, verifying a claim, controlling a workflow — these tasks need a label or a score, not a paragraph. Microsoft-Decision-1 is purpose-built for that class of work.

The model lives in Microsoft Foundry and is also reachable through OpenRouter.

That means developers can call it from existing applications, agents, and workflows without leaving the environments they already trust. The blog post emphasizes a secure, trusted environment, but it does not detail what specific compliance certifications or data-handling guarantees accompany the service.

How Microsoft-Decision-1 performs on benchmarks

The company ran a 36-benchmark comparison covering nearly 150,000 questions. The benchmarks were kept blind from training. Microsoft-Decision-1 achieved the highest average accuracy across the suite. On the JevBench leaderboard dated 8 October 2026 it holds the top position. The same source notes that the model is first on accuracy and the fastest measured.

Latency numbers are striking. The model runs 4.5 times quicker than Quyet-1.0-Large, the runner-up. It runs 35 times quicker than GPT-6 Sol, the reference LLM in the test. Hover or tap a row in the interactive chart for the detail.

Where you can use it and what it changes for developers

If you currently route requests with an LLM, you pay for tokens you don't need and wait for generation you don't read. A decision model returns a class ID or a priority score in a single forward pass. The announcement lists routing, classification, prioritization, verification, and workflow control as primary use cases. Each of these maps to a branch in code that previously required prompt engineering, output parsing, and retry logic.

The post does not say how many tokens the model consumes per decision, what the pricing tiers look like, or whether fine-tuning on private data is supported. It also does not disclose how many instances of the model are already deployed in production inside Microsoft products. What it does say is that the model is available today in Foundry and OpenRouter, and that it outperforms both LLMs and other decision models on the benchmarks the company chose to publish.