It didn't seem that the mid-end was efficient when I last went processor shopping (last June). It appeared that of the Zen 3 processors (which were interesting as I already had an AM4 rest-of-the-computer) only the 5600 and 5900X had "good" prices to me, as the 6-core 5700 was only a bit faster than the 5600 for a fairly large jump in price, and the 8-core 5800 had only two more cores than the former processors, while not being _that_ much cheaper than the 12-core 5900X. Given that I hold onto parts for a while, and that moar cores with shared memory is what I was after (for fuzzing and model checking, mostly[1]) I took the plunge and got the 5900X.
Another advantage there was that I would not have to build another computer to use two cheaper processors. But that only works as I don't think I am limited by any other factors like memory or disk bandwidth.
I did however limit the processor to a 65W TDP to reduce heat output and power usage, which only slowed down my very parallel programs about 10%, and didn't affect less parallel programs at all. (The duplicated other components of two machines would also draw more power, though perhaps not that much in comparison to processors here.)
[1] For a fuzzer you could easily give separate machines multiple random seeds, so really that is no issue. But for model checking/exhaustive search there needs to be a proper form of work distribution. I haven't a clue what the network bandwidth requirement of e.g. TLC, so that might also require faster NICs than one gets on motherboards to not stall processors on the network.