OCCT is a more stringent test than Realbench. The voltage difference you needed agrees fairly well with my experience. Cinebench is not as good as Realbench. My current 24/7 CPU passed a long Realbench test, but crashed at a critical point of a Creative Suite workflow. No such crashes since my stability criterion has been OCCT.
Lately I've been using y-Cruncher. Look it up on HWBOT. It chooses from several algorithms to use the set of AVX instructions in the CPU under test. It also needs very good memory stability. Most important, it checks the computations.
Prime 95 is more harmful than useful. Prime 95 is all stress and no test in that calculation results are not checked. When passing Prime 95 you know only that it did not crash.