Microsoft released Decision-1, a fast model for routing and classification

A team that pays a large model to pick between options on every step of an agent has a cheaper option today.

Microsoft published Microsoft-Decision-1, a model built only for structured decisions, and said it is available in Microsoft Foundry and coming to OpenRouter. Achint Srivastava, a vice president in the Office of the CTO, wrote the announcement. The model is for routing, classification, prioritisation, verification, workflow control and rubric-based grading of what another model has said. Each call returns a calibrated probability for each option on offer.

What the model is for

The input is a fixed set of answer options, in a yes-and-no, multiple-choice or rating shape, and the output is a probability score for each. That means an agent can ask it which tool to call next, which support ticket is urgent, whether a draft reply breaks a rule, or how to rate another model's answer against a rubric. Microsoft says it post-trained Qwen3.5-9B for single-pass decision scoring and will rebase the model on others later, including its own MAI and OpenAI models.

The latency and accuracy claims

Microsoft says Decision-1's P50 latency is about 35 times faster than GPT-6 Sol and 4.5 times faster than Quyet-1.0-Large, the runner-up in its own test. The company compared it against several top public models from the JevBench open leaderboard across 36 extra public and private benchmarks, nearly 150,000 questions total, kept blind from training, and reported the highest accuracy on that set.

Robustness and calibration

Microsoft says it also measured how often the decision flips when the same request is paraphrased or the options are reordered. Decision-1 changes its decision on 1.3 percent of eight-way perturbations on average, and zero flips when option descriptions are paraphrased or when options are reversed or shuffled. The probability itself is part of the API, which Microsoft says lets an application decide when to act on a prediction, defer, or send it to a human.

Every one of these figures is Microsoft's own, from its own benchmark run, so a team picking Decision-1 for its own workflow should still check the model against the exact decision it will actually make. The company's framing of the model as a System One tool next to a reasoning model matches the shape of several products that shipped in the past month, from TypeSafe Jev to Cloudflare's Clef and Ollaya's local runtime. The crowded market is a reason to read a vendor's own benchmark carefully rather than evidence that the problem is solved.

The deeper point is that a step in an agent that only needs one choice does not need a model that writes paragraphs. On a long workflow the saved tokens and the saved seconds add up: Microsoft's own example is a chain of 20 decisions where 100 milliseconds each adds two seconds to the end-to-end run.

Source

Microsoft-Decision-1: Our model for fast decision-making, Microsoft Command Line.