Based on how same models rank fluctuates week to week, all I can conclude is that no frontier models is statistically better than the other or it's too task dependent that the result cannot converge.

I will be trying these and will report back!