Microsoft launches Decision-1, a purpose-built model for fast classification and routing

Microsoft announces Decision-1

Microsoft has introduced Microsoft-Decision-1, an AI model built for a narrow job: making quick, structured decisions rather than generating text. In an announcement on the company's Command Line blog, which surfaced on Hacker News's front page on 9 October, Microsoft presents the model as a fast scorer for decisions, available now in Microsoft Foundry and due to reach OpenRouter in the near future.

The model is aimed at ordinary but high-volume software work: routing requests, classifying content, prioritising items, verifying outputs and steering workflows. The idea is that existing applications, agents and pipelines can hand these judgement calls to a dedicated model instead of improvising them with a general-purpose one, with Microsoft emphasising that this happens inside a controlled, trusted environment.

How decision models differ from LLMs

Microsoft frames decision models as a fast-emerging AI category in their own right. A large language model is built to produce prose or work through complicated reasoning; a decision model instead returns answers in a form software can consume directly — pick a label, return a score, choose a route. Because the task is tightly scoped, the argument goes, classification and decision-making can happen at very low cost and high performance, which unlocks jobs that would be uneconomical to run through a general model.

This is a familiar pattern in machine learning: narrow, specialised models have often beaten larger general ones on specific, high-frequency tasks. What is new is a major vendor packaging the idea as a distinct product category with its own name, distribution and benchmark story.

The performance claims

All of the numbers come from Microsoft's own testing and have not been independently verified. The company says Decision-1 recorded the highest accuracy across a comparison spanning 36 benchmarks and almost 150,000 questions, and that those benchmark sets were withheld from training so the model could not simply memorise the answers.

On speed, Microsoft reports that Decision-1 was the fastest model it measured: 4.5 times quicker than Quyet-1.0-Large, the runner-up in the comparison, and 35 times quicker than GPT-6 Sol. The company's headline claim is that the model leads on both latency and quality for structured decision tasks, ahead of general LLMs and rival decision models alike.

Availability

Decision-1 can be used today through Microsoft Foundry, the company's platform for building and deploying AI applications, with OpenRouter — a service that offers many models behind a single API — listed as coming soon. The distribution choice signals the intended audience: developers wiring decision logic into production systems, not chat users. Reaching OpenRouter also means the model will be reachable outside Microsoft's own tooling, where it can be compared directly against alternatives.

Why it matters

The launch matters less as a single product than as evidence of a category forming. For most of the AI boom, the default answer to any judgement call in a pipeline has been a general-purpose LLM: is this ticket urgent, does this request belong to billing or support, is this output acceptable. Those calls are high-volume, latency-sensitive and dull — exactly where a large reasoning model is the wrong tool. A cheap, fast scorer that returns structured answers changes the economics of agent frameworks, moderation systems and request routing, and Microsoft shipping one — backed by Foundry integration and OpenRouter distribution — suggests it expects the category to stick.

There are caveats. The benchmarks are vendor-run, the comparison set is described only in broad totals, and specialised classifiers live or die on edge cases that averages can hide. Whether Decision-1 holds up under third-party evaluation will determine whether "decision model" becomes a standard line item in AI architecture, or just a new label for work small classifiers already did.