Microsoft Decision-1: A Decision-Scoring Model Built for Agentic Routing

A Model That Only Makes Decisions

LLMs are becoming increasingly capable, but when building agents and automated workflows, developers often find they don't actually need an all-purpose generalist model. The most frequent needs are much simpler: pick one option from a fixed set, classify a request, verify whether a result holds, or decide which branch an agent should take next. These tasks are fundamentally decisions, not long-form generation.

Microsoft-Decision-1, launched on Product Hunt, is a model built specifically for this use case. It's defined as a "decision-scoring model" focused on four task types: routing, classification, verification, and agent control.

How It Actually Works

Unlike asking a general-purpose LLM to "think it over and then respond," Decision-1 works in a much more direct way: you give it a set of fixed candidate options, and it returns a calibrated probability for each option in a single pass.

Two key concepts here are worth unpacking.

Calibrated Probabilities

Even when a general-purpose LLM can make a choice, it struggles to provide reliable confidence scores — its "certainty" often has little to do with actual accuracy. Decision-1 emphasizes that its outputs are calibrated probabilities, meaning when it assigns 80% confidence to an option, that number is statistically closer to the true correct rate. For workflows that need to set thresholds or manage risk, this matters enormously.

Calibration is a statistical concept in probabilistic prediction that measures how well a model's expressed confidence aligns with its actual accuracy. A perfectly calibrated model means that among all predictions made with 70% confidence, the true accuracy should also be close to 70%. Most general-purpose LLMs suffer from overconfidence — even when their outputs look like probabilities, they reflect token distribution preferences more than genuine uncertainty estimates. Specialized post-training on decision tasks with a calibration loss in the training objective is a common technical approach to producing "trustworthy confidence scores." For automated workflows that rely on decision thresholds (e.g., "route to human review when confidence falls below 60%"), the statistical reliability of confidence scores directly determines the system's real-world usability.

Structured Output

Decision-1 returns structured output that software can consume directly without additional parsing. This eliminates the fragile pipeline of "get model to output text → extract result with regex or a secondary model," reducing engineering complexity and the risk of errors downstream.

Speed and Cost Are the Main Selling Points

According to Microsoft's official figures, Decision-1 was post-trained from Qwen3.5-9B, keeping its parameter count relatively lean. Microsoft claims it is approximately 35× faster than GPT-6 Sol in median latency.

For decision-type tasks, speed is nearly everything. A complex agentic pipeline might require dozens of consecutive routing and verification judgments — calling a heavyweight general-purpose model at each step would result in substantial cumulative latency and cost. A fast, cheap "decision node" model can dramatically improve the overall system's responsiveness.

On pricing, Decision-1 is now live on Microsoft Foundry at $0.042 per million input tokens, with output tokens free. Since decision tasks typically involve longer input context but extremely short outputs (often just a probability distribution over options), the "free output" pricing is a natural fit for the use case.

Post-training refers to further training applied on top of an already pre-trained base model, targeting specific tasks or capability goals — typically including supervised fine-tuning (SFT) and reinforcement learning alignment (e.g., RLHF). Choosing Qwen3.5-9B as the base rather than training from scratch reflects both efficiency considerations — a 9B parameter model already has sufficient language understanding for classification and routing tasks — and the fact that it inherits Qwen's foundational multilingual and instruction-following capabilities. Starting from an open-source base model and applying task-specific post-training is the mainstream approach to building vertical-domain models today: it compresses training costs while using specialized data and objective functions to significantly improve performance and efficiency on the target use case.

Why Specialized Models Deserve Attention

The emergence of Decision-1 reflects an increasingly clear trend in AI engineering: use the right model for the right job.

Over the past two years, the industry defaulted to throwing the most powerful general-purpose model at every problem — but in production, that often means high costs and unnecessary latency. As agents and automated workflows scale to production deployments, developers are increasingly breaking systems into specialized components: large models for reasoning and generation, lightweight specialized models for scheduling, routing, and verification.

Decision-1 is essentially playing the role of a "traffic cop" or "flow splitter" within a workflow. It doesn't need to generate impressive answers — it just needs to make fast, accurate judgments and return confidence scores that programs can use directly. This division of labor is precisely the kind of infrastructure needed to build stable, economical, and controllable agentic systems.

A Few Caveats

As a newly released product, public information on Decision-1 remains limited. Microsoft's claim of being "35× faster than GPT-6 Sol" is self-reported and lacks independent third-party benchmark validation — real-world performance remains to be seen. Additionally, the maximum number of candidate options supported, the range of languages covered, and accuracy on complex multi-layer decision tasks all require more real-world use cases to properly evaluate.

That said, from a product positioning standpoint, Decision-1 is targeting a genuine and fast-growing need. For developers building agentic systems who are frustrated by general-purpose models being both slow and expensive, it offers a new option worth exploring.