Microsoft Unveils Decision-1, a High-Speed AI Model for Agent Routing and Fast Decisions
Microsoft has launched Decision-1, a specialized model based on Qwen3.5-9B that delivers 83.5% benchmark accuracy with an 85 ms latency for rapid evaluation, classification, and agent control.
Microsoft has entered the rapidly evolving decision model race with the release of Decision-1, an AI model engineered specifically for high-speed, structured choices. Designed to handle classifications, evaluations, and routing tasks, the model also offers the capability to guide and control autonomous AI agents navigating complex digital environments.
The launch places Microsoft alongside a growing cohort of providers developing low-latency routing and decision engines, including OpenAI, Cloudflare, and startup Jev.
Architecture and Benchmark Performance
Decision-1 is built upon the open-source Qwen3.5-9B architecture. According to Microsoft, the system led performance evaluations across 36 distinct benchmarks covering nearly 150,000 questions.
In internal testing against competing lightweight engines, Decision-1 achieved an accuracy rate of 83.5% with an average latency of 85 milliseconds, outperforming Jev 1.13.0. Microsoft also reported that Decision-1 is 2.5 times faster than the runner-up in speed, H2O-Lightning-4B. Notably, Cloudflare’s open-source Clef models—which are also based on Qwen—were not included in Microsoft's benchmark comparison.
Pricing and Access
Microsoft is making Decision-1 accessible to developers via Microsoft Foundry and the third-party model aggregator OpenRouter.
To encourage adoption in high-frequency routing pipelines, the model features an aggressive pricing structure:
- Input tokens: $0.042 per million
- Output tokens: Free
The Rapid Shift in Decision Models
The decision model segment has seen accelerated disruption within a short timeframe. Jev initially triggered the trend in mid-September by demonstrating the value of dedicated, ultra-fast routing systems over general-purpose large language models.
However, open small language models have enabled rapid adaptation across the industry. With larger players fine-tuning compact base models like Qwen into high-accuracy, sub-100-millisecond decision engines, the infrastructure layer for orchestrating multi-agent systems is becoming increasingly competitive and cost-effective.
