On October 9, Microsoft introduced Microsoft-Decision-1, a decision-scoring model now available on Microsoft Foundry and OpenRouter. Rather than generating text paragraphs, the model returns probabilities for each option within a predefined set of choices. Microsoft is targeting the product at data classification, model routing, result verification, and workflow steering for AI agents.
This approach isolates part of the work typically assigned to LLMs into a narrower task: deciding which step should be executed next. What is notable is not conversational ability, but whether a fast scoring layer can help software reduce latency while still selecting the correct action.
Scoring choices instead of writing answers
Microsoft-Decision-1 was developed by fine-tuning Qwen3.5-9B for single-pass decision scoring. The model supports yes/no choices, multiple-choice questions, rating scales, and rubric-based response quality scoring. Microsoft stated it will continue building products on other models, including Microsoft AI and OpenAI; this remains a plan, not the current configuration.
Probabilities are part of the API output. Applications can use confidence levels to select an action, defer it, or escalate to a human reviewer. However, confidence figures are only useful when properly calibrated on data representative of real-world tasks. A high-probability result does not in itself guarantee that the suggested action is safe.
High speed on decision tasks
According to the company, Decision-1 achieved the highest accuracy in a comparison across 36 benchmarks with nearly 150,000 questions, including evaluation sets not used during training. Microsoft says its median latency is roughly 35 times faster than GPT-6 Sol, with measured throughput 4.5 times faster than Quyet-1.0-Large. These are self-reported results from Microsoft without independent verification in the referenced source; they cannot be extrapolated to imply that it can replace GPT-6 Sol across all workloads.
The company also tested stability using eight input perturbation types, such as rephrasing prompts or shuffling choice order. The reported average rate of decision changes was 1.3%. This result describes Microsoft’s test and is not a guarantee that the model will always make the same decision in every situation.
In internal testing, Xbox Research used the model to classify over 10,000 feedback responses into predefined topics. Microsoft reported that quality was competitive with GPT-6 Sol, with 14 times faster speed and 200 times lower costs. Those ratios pertain to that specific Xbox team task, not a default level of savings across all applications.
Low pricing does not replace agent permission guardrails
The price announced by Microsoft is $0.042 per million input tokens, with output tokens being free. This pricing aligns with building a decision layer invoked multiple times throughout a pipeline, rather than relying on a large model to handle every step. Overall system costs still depend on invocation frequency, input volume, and the downstream tools called after each decision.
Practically speaking, the real value lies in proper separation of responsibilities: the scoring model chooses among options, while the execution system enforces permission boundaries and requires confirmation for sensitive actions. As Decision-1 is adopted beyond Microsoft’s trials, what bears watching is whether its accuracy and reliability hold up on application-specific data.