Running the decision against real prompts through OpenRouter instead of synthetic ones is the part most MoE-vs-dense writeups skip, and the three-way recommendation (use / marginal / stick) matches how I actually want to read a benchmark. One thing I'd add on top of latency + cost per 1k: a quality-delta gate against a golden set, because a MoE that beats dense on both SLAs but drops 4 points of task accuracy is a bad switch dressed up as a good one. Did you consider stapling an eval score into the decision matrix?