Most AI costs in a real workflow don’t come from the big expensive model. They come from the dozens of small questions your system needs to answer before the big model ever speaks: Is this request safe to process? Which team should handle it? Does this response meet quality standards? Those questions get routed to large language models by default, and that’s where costs balloon.
Microsoft just released a model called Microsoft-Decision-1, and it was built specifically for those smaller, faster decisions. Decision models are a new and distinct category from language models. Where a language model generates text word by word, a decision model takes a fixed set of options and returns a calibrated probability score for each one. That’s it. One structured output, immediately usable by your software.
The numbers are hard to ignore. Microsoft-Decision-1 runs 35 times faster than GPT-6 Sol on structured decision tasks and costs $0.042 per million input tokens, with output tokens completely free. Teams at Microsoft have already used it to process over 10,000 pieces of customer feedback 14 times faster and 200 times cheaper than GPT-6 Sol with comparable accuracy. Xbox Research used this to categorize game reviews. The Copilot team used it for quality control on chat responses. On-call engineers used it to route incident response data across logs and ticketing systems.
The model is built on a fine-tuned version of Qwen3.5-9B, a 9-billion parameter base optimized for single-pass decision scoring. Nine billion parameters sounds technical — think of it as a smaller, focused engine built to answer one type of question instead of a general-purpose model that can do everything. It supports yes/no, multiple-choice, and rating options, plus rubric-based grading of AI responses and agent actions. You send it a question and a list of possible answers. It tells you which one fits best, along with how confident it is.
For businesses building AI-powered workflows, this matters. Every time your system needs to route a customer request, classify a document, validate an input, or check whether an AI-generated response is acceptable before sending it, you can now do that job at a fraction of the cost. If you have a workflow with 20 sequential decisions and each one saves 100 milliseconds, you’ve just cut two full seconds off your response time. That’s the difference between a product that feels instant and one that feels slow.
The broader trend here is worth watching. AI is splitting into two tracks: generative models for complex reasoning and creative tasks, and decision models for structured, high-frequency classification work. Trying to use GPT-level models for every decision in your pipeline is like hiring a senior architect to sort your mail. The right tool for the job matters more as AI workloads scale. Microsoft-Decision-1 is available today in Microsoft Foundry and through OpenRouter.
Want to explore how decision models could benefit your business? Let’s talk.