Did you account for all the vulnerabilities paches in Intel? Some benchmarks might now show it, but in real life you'll get that performance hit (or you'll need to disable those protections and some people don't like it).
9600k, 9700k, and 9900k are all single threaded king, effectively, at least until Zen 2 launches on 7/7.
9600k has a 3.7ghz stock base clock, 9700k and 9900k have identical base clocks of 3.6; in order, their stock single core turbos are 4.6, 4.9, and 5.0.
Now, you could be asking yourself "but Diablo, in single threaded, the 9900k is 8.6% faster, and the 9700k is 6.5% faster than the 9600k, respectively". They're all k parts, just overclock it the tiny bit of the way for a fraction of the price ($230 vs $410 vs $490), and the 9600k has far more thermal and power budget per-core to play with than the other two do.
I almost built a new desktop with a 9600k, the $230 6/6 part that shames (in both single and multithreaded) the highest end $340 or $350 4/8 parts from the previous four generations (4770k, 4790k, 6700k, 7700k); vs the $360 8700k (6/12), it performs identically in single threaded, and illustrates that hyperthreading only gains you about 25% extra performance if you can saturate all 12 threads.
Remember, these are desktop chips, you are very unlikely to ever need more than 8 threads (and use them effectively), even if you game. If you happen to have a use case where you can easily saturate >8 threads, anything LGA115x is probably inappropriate for you anyways.
However, with the 9000 series release, Intel admitted AMD scared the fuck out of them, and re-released Coffee Lake at a lower price with some tweaks in speed and core/thread count because they were afraid of Zen 2 being a success; their fears were warranted.
The $250 3600x (a 6/12 part) beats the 9600k in single threaded, widely beats it in multithreaded, and has PCI-E 4.0, with an extra 4 CPU-bound PCI-E lanes (20 vs 16 on AM4 vs LGA115x), and uses a higher DDR4 clock (stock 3200 vs 2666, with an effective maximum of possibly over 5000 vs around 4133; AMD's IMC seems to continue to be effective at lower CAS latency than Intel's is).
Side note: 4790k, a Devil's Canyon part, is Intel's stand-in for the non-existent 5700k.
Devil's Canyon is a Haswell Refresh part that got a second set of 4000 series model numbers instead of 5000 series. Haswell Refresh was a Broadwell core paired with a DDR3 controller, fabbed at Haswell's node size; there are no architectural changes between Broadwell and Haswell in the core, only major changes was the decrease in node size and the swap to DDR4, the core design remained nearly identical. Broadwell largely ignored the desktop, focusing on LGA2011 and mobile parts instead, leaving Haswell Refresh to fill in the gap with the desktop and Xeon E3s, and the only notable exceptions being a small set of Iris Pro GPU parts.
Sorry, I should have written major changes. In practice, benchmarking identical Haswell vs Haswell Refresh (that, again, are effectively DDR3 Broadwells), such as with an E3-1230v3 vs a E3-1231v3, I did not see anything that wasn't within a reasonable margin of error.
At Computex they were claiming the 4.5Ghz 3800X has 1% higher single threaded performance in Cinebench than the 9900K, presumably both were measured at stock clocks. The 3950X has 4.7Ghz boost clock stock so it'll be interesting to see how high ST performance can be pushed with overclocking.