This is an unfair comparison, the incentive for the close sourced models are way too high for their models to behave and follow the prompts. The amount of resources put by the proprietary model is no way near these open source models.
I also can design tests. A test with my local model got 0% downloads from malware command control center while closed models download malware.
We design the test we want, to make what we want to look good.
Open-weight models 100% more inspectable for all sorts of things, transparency trumps