Microsoft Launches Decision-1 Model for Automated AI Classification Tasks
Microsoft releases Decision-1, an AI model designed to return probability scores for predefined answers, offering lower latency and cost compared to general large language models.
Key points
- Microsoft Decision-1 returns calibrated probabilities for fixed answer sets rather than generating free-form text.
- The model is built on Qwen3.5-9B and is available via Microsoft Foundry and OpenRouter platforms.
- Microsoft claims the tool is significantly faster and cheaper than general-purpose LLMs for specific decision tasks.
Microsoft has officially launched Decision-1, a new artificial intelligence model designed to separate decision-making logic from generative content creation. The system returns probability scores for predefined answer options, allowing software applications to process outputs directly without interpreting free-form natural language responses from large language models.
The company states that this approach reduces both computational costs and processing latency for tasks requiring specific judgments. According to Microsoft, the model handles operations like classifying requests, prioritizing incidents, and verifying AI-generated responses more efficiently than general-purpose systems.
Technical Specifications and Performance
Microsoft constructed Decision-1 by post-training the Qwen3.5-9B model. The firm plans to rebase the technology on other architectures in the future, including Microsoft AI and OpenAI models. The tool is currently accessible through Microsoft Foundry and OpenRouter, providing developers with immediate integration options for their workflows.
Decision-1 supports binary judgments, choices among predefined options, ratings, and rubric-based assessments of agent actions. Unlike standard large language models that generate explanatory text, this model returns probabilities for available answers in a single pass. Microsoft noted that the system can also evaluate multiple decisions in parallel within one request, streamlining complex agent workflows.
Benchmark Claims and Pricing
Microsoft reported that Decision-1 achieved the highest accuracy across 36 benchmarks covering nearly 150,000 questions that were kept blind from training. The company claimed the model is approximately 2.5 times faster than the second-place competitor and 35 times faster than GPT-6 Sol. These performance metrics remain unverified by independent third parties.
Pricing for the service is set at 4.2 cents per million input tokens, with no charges for output tokens. In an incident-prioritization example, Microsoft calculated that a request using 86 input tokens and two output tokens would cost roughly $0.0000036. At this rate, one dollar would cover more than 275,000 similar decisions, according to the company’s blog post.
Industry Reaction
Charlie Dai, vice president and principal analyst at Forrester, commented that specialized decision models could significantly reduce the cost and latency of agentic AI workflows. This shift suggests a move toward more granular AI architectures, where specific tasks are handled by optimized models rather than relying on broad, general-purpose systems for every operation.
What should you do?
Organizations should evaluate their current AI workflows to identify tasks that involve repetitive classification or prioritization. Security teams can monitor how these specialized models are integrated into critical decision-making processes to ensure accountability. Developers should review the latency and cost benefits for high-volume decision tasks.
Background: Large language models
Large language models are a specific type of neural network trained on text. Generative AI is the broader umbrella term for any system that creates new content, including images and code. Not all generative AI uses large language models, but most modern text generators do.
What to do and how to stay safe: Microsoft
- Audit existing AI workflows to identify tasks that rely on general-purpose models for simple binary or categorical decisions.
- Evaluate the accuracy requirements of your decision-making processes to determine if a probability-based model meets your quality standards.
- Monitor vendor documentation for updates on model rebasing and integration capabilities with your current infrastructure.
- Review cost structures for high-volume API calls to assess potential savings from switching to specialized decision models.
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
What is Microsoft Decision-1?
Decision-1 is an AI model that returns probability scores for predefined answers, designed for tasks like classification and prioritization rather than free-form text generation.
How is Decision-1 different from a standard large language model?
Unlike standard models that generate natural language explanations, Decision-1 provides calibrated probabilities for fixed answer options in a single pass, reducing latency and cost.
Is there an independent verification of the performance claims?
No, the benchmark results and cost calculations are Microsoft’s claims and have not been independently verified by external parties.
