This is mostly in fact an example of a big problem with benchmarks. What's measured here is the difference between never using an object (and so the actual work to prepare it can be elided, albeit by the OS not the programming language) and doing all the work up front in the expectation you'll use it, but then not using it.

u8 turns into "Hey, Linux kernel, zero these pages if I ever read them" (and then never reading them) whereas the opaque type turns into memset() zeroing all the pages.

Not doing any work is in fact a million times faster but your real program would have needed to do work, otherwise why bother having the vector? Whereupon the benefit disappears.

Where possible design your benchmarks to really do the thing you think you're measuring. If what you're measuring is nothing then be sceptical about supposed "performance" measured for that, since it's nothing, you're probably exploring the same space as the people who wanted to find out how much the human soul weighs (trick question, there is no such thing, but they put a lot of effort into trying to measure it anyway).