Microsoft launches Microsoft-Decision-1 model

Microsoft post-trained Alibaba's Qwen3.5-9B to create Microsoft-Decision-1, a model that delivers calibrated probabilities for structured choices instead of unformatted text. Available on Microsoft Foundry and OpenRouter via Azure, the system features a 32,768-token context window and produces JSON outputs. Microsoft charges $0.042 per million input tokens, while output tokens are free. The model is hosted strictly via proprietary API without open weights or downloadable quantized versions.

In vendor-run evaluations across 36 blind benchmarks comprising 147,137 questions, Microsoft-Decision-1 achieved an industry-leading 83.5% average accuracy. It recorded a p50 latency of 85 ms and p95 of 125 ms, outperforming Quyet-1.0-Large, which achieved 81.9% accuracy with a 380 ms median latency. GPT-6 Luna Decisions scored 79.4% accuracy at 300 ms latency and $0.10 per million input tokens, while H2O-Lightning-4B scored 77.2% accuracy. Compared to the 3.01-second execution of GPT-6 Sol, Microsoft-Decision-1 ran 35 times faster. On a 92.2 calibration score, it trailed only Quyet-1.0-Large's 93.1 rating.

Stress tests involving eight request perturbations showed the system changed decisions on just 1.3% of variants, maintaining perfect consistency when options were shuffled or paraphrased. Safety evaluations spanning 5,250 queries across 11 benchmarks verified prompt injection and jailbreak defenses. Internally, Xbox Research used the tool to analyze over 10,000 feedback entries 14 times faster and 200 times cheaper than GPT-6 Sol. Microsoft Discovery noted 46 times greater scoring consistency than standard LLMs, while Copilot quality control ran 100 times faster than GPT5.6 Luna. For software engineers, the model offers a high-speed routing and verification engine, though latency metrics remain contested by rivals like H2O.ai, whose H2O-Lightning-4B ranks first on official JevBench leaderboards.