I'm not sure it got mentioned yet, but there's been an upload of Winstone 95 since the 686 benchmark results were run (an ISO of it was posted in the Vogons Archive as of 2016) and the Ziff Davis Winbench 95 software package was posted on PCCorner in the last few years, too.
And given those are the tests that predominantly championed Cyrix 6x86 performance, it's interesting to look back at those. I've only run tests with my Asus P5A-B so far (which isn't the fastest for 6x86s/MIIs) but it does seem to back up the relative performance. And I also started doing comparisons with P54 and P55 CPUs at equal clock and bus speeds for more 1:1 comparison (overclocked bus Pentiums is something Redhill didn't touch on). The bit I run, including 2x75 for the 6x86 and P54 still left the 6x86 well ahead, though.
Using the 1x multiplier setting also gets some interesting results that favor some of the Ziff Davis tests and points to some interesting in-era possibilities for early adopters with the 80 or 100 MHz rated parts. (dropping down to 1x 66, 68, or 75 MHz getting a performance boost for a good number of things ... 83 MHz moreso if you had a board capable of it back then, or maybe upgraded to a new board+RAM with an old CPU in 1997)
They don't seem to like 1x100 MHz, though. (the one 80 MHz, PR90 CPU I have still seems to do 1x 83 MHz fine in Win98SE)
I also have an FIC VA-502 now, so I can test Red Hill's claims of unusually good performance with 6x86MX (and presumably MII) chips, particularly their PR200 (166 MHz) configuration, which was supposed to be the fastest general-purpose build they could make at the time.
The only other area that's not been fully explored in retrospect (I think) is the relative gaming performance of Cyrix builds for things other than Quake and most API-driven 3D engines, but ones that were still pretty CPU demanding at the time, even ones a few years old (like 1993-1995 era 3D games with 640x480 SVGA rendering support).
Phill's standard Socket 7 benchmark comparison videos seem to include Wind Commander III's hanger launch sequence for minimum framerates, and is one interesting area to start. He also seems to have found out that faster clocked MX/MII CPUs run too fast for Wing Commander 1 even with all cache disabled, so you need to drop the multiplier or bus speeds down for that (I believe it's too fast by the PR300 range, so 233 MHz).
OTOH, with an earlier S7 board capable of 50 MHz FSB, 1x50 MHz with a 6x86 might work for Wing Commander II without disabling the L1 cache.
There's also the odd case of Tomb Raider, at least for its software renderer, that seems to be fairly heavy on the ALU side and doesn't use a Quake-like FPU-driven pixel pipeline, but does still require an FPU. It might be optimized for 486 DX4s and x5s or something like that, and seems like it's well balanced for the beefy (for 486) FPUs in Cyrix 486s and 5x86s as well as AMD's overclockable 5x86. (I think the AM5x86 also might have a slower ALU than the Cyrix or Intel DX4, or enough that offloading mul/div heavy math to the FPU might even be faster ... though that may be true of most/all 486s if you need 32-bit precision or better: 16-bit integer math has more speed advantages on the 486)
That sort of balance might also carry over well to the Cyrix 6x86 and K5 with their balance of ALU and FPU performance.
Other than Tomb Raider, I'd think most other relevant mid-90s games that might be faster (at least per-clock, if not by PR-number) would be ones released prior to 1996, or at least using engines designed well prior to that. (like late generation Doom or Build engine games, or Rise of the Triad's 32-bit extended Wolf3D engine)
And I'm not positive, but the DOS, CD-ROM version of Tie Fighter might be the most taxing 16-bit real mode game that could be tested, with its 640x480 SVGA mode. (I haven't confirmed that's a real-mode game, but given the engine it's built on and the fact it still requires conventional + EMS memory exclusively, seems to imply that pretty heavily) It's on my list of things to get running on a 286. (I would have already done it if I hadn't been dragging me feet with setting up a CD-ROM driver)
I'm not sure what it takes to max out Tie Fighter (especially in more populated/active scenes) but it might still be relevant for even a Pentium 100-166 class system. And the 6x86 is known for being very fast at 16-bit real-mode operations (and presumably in its Virtual 8086 implementation) so would be a good contender there. Also potentially one of the few areas outside of 3DBench that the 6x86 would vastly outperform the Pentium Pro, or somewhat outperform it at the same PR number.
OTOH, the X-Wing and Tie Fighter games are somewhat well known for their EMS-handler related errors (EMM386/XMS errors), and the original Floppy Disk version of X-Wing is particularly crash prone and speed-sensitive (or video card sensitive) in my experience on a SS7 system (the Tie Fighter and related BWing/XwingCD engines are more stable, but still EMS driver sensitive). So they'd be good tests for stability and, unlike Wing Commander 1 and 2, they're properly frame limited (using some sort of timer interrupt or polling routine) and not CPU-clock dependent, so can be maxed out on fast systems and kept playable.