Microsoft Launches Decision-1 AI Model Built On Qwen3.5-9B For Faster, Cheaper Decisions

Microsoft has launched Microsoft-Decision-1, a new artificial intelligence (AI) model designed to make fast, low-cost decisions across enterprise applications and AI agents. Built on Alibaba's Qwen3.5-9B model, it scores predefined choices instead of generating text, with Microsoft claiming significant improvements in decision-making speed and accuracy.

TL;DR

 

• Microsoft introduced Microsoft-Decision-1 on October 9, 2026, making it available through Microsoft Foundry and OpenRouter.
• The model uses Alibaba's Qwen3.5-9B architecture for classification, routing, verification, and AI agent control.
• Microsoft claims it was 35 times faster than GPT-6 Sol in its tests, with pricing starting at $0.042 per million input tokens.
• Its performance claims are based on Microsoft-led evaluations.

According to an official Microsoft announcement, Microsoft-Decision-1 is designed to handle structured decision-making tasks that would otherwise require larger AI models.

Source

Unlike conventional large language models (LLMs), which generate responses or reason through complex problems, Microsoft-Decision-1 evaluates predefined options and assigns a probability score to each one. This allows applications to classify information, verify responses, prioritize tasks, and determine the next action in automated workflows.

“Decision models are quickly emerging as an important new category in AI,” said Achint Srivastava, VP of Software Engineering, Office of the CTO at Microsoft, in the announcement.

Microsoft developed the model by post-training Alibaba's open-weight Qwen3.5-9B architecture for single-pass decision scoring. The company also plans to use Microsoft AI (MAI) and OpenAI models as the foundations for future versions.

As highlighted by MarkTechPost, the model supports yes/no questions, multiple-choice responses, ratings, and rubric-based assessments of AI-generated responses and agent actions.

It also features a 32,768-token context window, text-only inputs, and structured JSON outputs.

Microsoft Claims Higher Accuracy And Faster Decision-Making

Microsoft says that Microsoft-Decision-1 achieved the highest accuracy in its evaluation of 36 benchmarks, involving nearly 150,000 questions that were kept separate from model training.

The company reported that its model was 2.5 times faster than H2O-Lightning-4B v1.1 and approximately 35 times faster than GPT-6 Sol in its benchmarking.

Meanwhile, MarkTechPost reported an average accuracy of 83.5% and median latency of 85 milliseconds, based on Microsoft's published results.

However, the publication also noted that the latency comparisons were not entirely equivalent, as Microsoft's model was measured through Foundry while competing models used adjusted benchmark timings.

Microsoft additionally tested the model's consistency by modifying requests in eight different ways, including changing option descriptions and rearranging choices. It reported that the model changed its decisions on just 1.3% of these variations on average.

On safety, the company evaluated the model against 5,250 requests across 11 benchmarks covering harmful content, jailbreak attempts, and prompt injection.

These results suggest potential improvements for structured AI workloads, although the performance comparisons remain vendor-reported rather than independently validated.

Microsoft Is Already Testing Decision-1 Across Xbox, Copilot And Other Teams

Microsoft also revealed that several of its internal teams have been experimenting with the new model.

Xbox Research used Microsoft-Decision-1 to categorize more than 10,000 pieces of player feedback and reviews into predefined themes. According to Microsoft, the model delivered quality competitive with GPT-6 Sol while operating more than 14 times faster and at 200 times lower cost.

The Microsoft Copilot team also tested the model for evaluating chat and agent responses, finding it competitive with GPT5.6 Luna while operating 100 times faster.

Meanwhile, Microsoft Discovery used the model to assess experimental outcomes and support adaptive replanning in scientific workflows. Microsoft reported that decision scoring was 46 times more consistent than its LLM-based approach and three times faster.

The company also identified applications in incident response, model routing, data validation, search relevance, content classification, and AI agent guardrails.

For developers building multi-step AI agents, these capabilities could reduce the need to call larger generative models for every routine decision.

Microsoft-Decision-1 Pricing And Availability

As confirmed by the Microsoft Foundry model catalog, Microsoft-Decision-1 is accessible through Microsoft's AI development platform.

It is also available through OpenRouter, allowing developers to integrate its decision-scoring capabilities into supported applications and workflows.

Microsoft has priced the model at $0.042 per million input tokens, with no charge for output tokens.

Dealroom News noted that the model is positioned as a low-cost control layer for AI agent workflows, helping applications make structured decisions without repeatedly invoking larger models.

However, Microsoft makes clear that the model is not designed for general conversation, summarization, translation, or open-ended question answering. It also does not generate explanations for its decisions.

Furthermore, the company's model documentation cautions against using Microsoft-Decision-1 as the sole automated decision-maker for consequential matters involving employment, credit, housing, healthcare, or legal rights.

With Microsoft-Decision-1, the company is expanding its AI portfolio beyond text generation and reasoning models to include specialized systems for structured decisions.

The launch could give developers another option for reducing costs and delays in AI-powered applications, although the model's effectiveness will ultimately depend on real-world workloads and further evaluation.