Microsoft has released Microsoft-Decision-1, a model that is not trying to be another general chatbot.

Its job is narrower: given a fixed set of options, return a structured decision with probability scores. That makes it a model for routing, classification, prioritization, verification, evaluation, and workflow control rather than open-ended generation.

That difference matters. A lot of production AI systems do not need another paragraph of prose at every step. They need to answer questions like: Which queue should this ticket go to? Is this output acceptable? Should this agent continue, retry, stop, or escalate? Which model is cheap enough and capable enough for the next step?

Decision-1 is Microsoft's answer for that layer.

What Microsoft-Decision-1 does

Microsoft says Decision-1 is available in Microsoft Foundry and through OpenRouter. Foundry Labs describes it as a decision model for fast, repeatable tasks in AI applications, especially cases where software needs to choose among predefined options and use confidence signals directly.

The model is built on Qwen3.5-9B and post-trained for fast, single-pass decision scoring. Instead of generating a free-form response, it evaluates the input against choices supplied by the developer and returns a probability for each choice.

Microsoft says it supports yes/no decisions, multiple-choice decisions, rating options, and rubric-based grading for AI responses and agent actions. Vercel, which also added the model to AI Gateway, describes the same pattern: ask typed questions about text and receive probabilities, choices, and scores.

That makes Decision-1 a useful fit for the unglamorous but constant decisions inside AI software:

  • Classifying support requests by department or urgency.
  • Routing an agent to the right tool or model.
  • Scoring whether an answer meets a rubric.
  • Deciding whether a workflow should continue or hand off to a human.
  • Validating whether an input meets a policy or formatting rule.

The important constraint is that Decision-1 is not meant to write the final answer. It is meant to choose among defined options.

Why Microsoft is positioning this as a new model category

The release is part of a broader shift from "one big model does everything" to modular AI systems.

In that architecture, a frontier language model might handle reasoning or writing, while smaller specialized models handle routing, moderation, scoring, classification, and guardrails. Microsoft is giving that decision layer its own model and API surface.

Microsoft says Decision-1 achieved the highest accuracy in its 36-benchmark comparison, covering nearly 150,000 questions that were kept blind from training. It also says the model was the fastest in that benchmark set, including a claimed 35x P50 latency advantage over GPT-6 Sol and 2.5x over H2O-Lightning-4B v1.1.

Those are Microsoft-run benchmarks, so teams should test on their own tasks before redesigning workflows around the claim. But the direction is practical: if a step only needs a label, route, score, or allow/block decision, using a large generative model can be wasteful.

Pricing reinforces that positioning. Microsoft lists input tokens at $0.042 per million, with output tokens free. OpenRouter lists the same price and a roughly 33K-token context window for the hosted model.

Where this fits in agent systems

Decision-1 is most interesting around agents, because agents constantly need control decisions.

An agent that can browse, call tools, write code, open tickets, or trigger workflows should not blindly continue just because a language model produced a plausible next step. It needs checkpoints: is this safe, useful, on task, and worth doing now?

Microsoft's examples include agent controls, model routing, AI judging, incident response routing, data validation, safety screening, computer and UI use, and scientific discovery workflows. Internally, Microsoft says teams have tested Decision-1 for Xbox feedback labeling, Copilot response quality checks, incident-response retrieval, and adaptive replanning in Microsoft Discovery.

That is the stronger story than the raw benchmark number. Decision-1 is a sign that big AI platforms are starting to productize the control plane around agents, not just the generative model at the center.

For builders, the question becomes: which parts of an AI workflow should generate, and which parts should decide?

What teams should watch

Decision models are only as useful as the choices and rubrics developers give them.

If a workflow asks vague questions, gives overlapping categories, or uses a bad escalation threshold, calibrated probabilities will not save it. The engineering work moves from prompt writing to option design, evaluation data, confidence thresholds, and fallback behavior.

That is still a good trade for many production systems. A structured decision is easier to log, test, compare, and audit than a paragraph of free-form text. It is also easier to put behind an approval gate: act automatically above one confidence threshold, ask for review below it, and block outright for risky categories.

The release also creates a new evaluation question for teams using AI agents. Instead of only asking "which model writes the best answer," teams should ask "which model should decide what happens next?"

Microsoft-Decision-1 is built for that second question.