Asif Razzaq writes that Microsoft released Microsoft-Decision-1, a model post-trained from Qwen3.5-9B that scores fixed answer options with calibrated probabilities rather than generating text. The model is designed for routing, classification, verification, and agent control, offering a 32,768-token context window, 85 ms p50 latency, and 83.5% average accuracy across 36 internal benchmarks. It is available as a hosted API through Microsoft Foundry and OpenRouter at $0.042 per million input tokens with free output tokens, but it has closed weights and text-only limitations.
- Latency comparisons against competitors are disputed because Microsoft used its own measurement while citing JevBench adjusted figures for others
- The model achieves zero decision flips under paraphrasing, reversal, or option-shuffling perturbations
- Internal Microsoft use cases report 14-100x speed advantages over GPT-6 Sol in feedback sorting and quality control
- Latency comparisons against competitors are disputed because Microsoft used its own measurement while citing JevBench adjusted figures for others
- The model achieves zero decision flips under paraphrasing, reversal, or option-shuffling perturbations
- Internal Microsoft use cases report 14-100x speed advantages over GPT-6 Sol in feedback sorting and quality control
