An article from 2000 on embedded.com dates the Hitachi 64180 (which Zilog second sourced as the Z180) to 1985, the Z280 to 1976, but that's almost certainly a typo for 1986, the Z380 to "the mid '90s", the Rabbit 2000 to "last fall" (i.e. 1999), and the eZ80 explicitly to 1999. (Aside: it's an interesting read beyond those data points, giving a lot of technical detail rather than merely a dry recitation of dates).
As per my comment, I'd add to that the R800, which was developed by the ASCII Corporation as an enhancement to the Z80 that had powered MSX machines until then; the bottom dropped out of the MSX market in the interim but it made it to market in the Panasonic Turbo R in 1990.
The 64180's product sheet states that it has an operating frequency up to 10Mhz. The same is true of the Z180. In both cases instructions have already become more efficient than the Z80.
The Z280 includes a clock divider and a clock output, the intention being that you feed it a fast clock and then it produces a divided one for everybody else and inserts wait states if and when necessary so that it isn't talking too quickly. It includes a cache, so a lot of the same considerations apply as with a more conventional clock multiplier, in addition to the instructions generally being more efficiently implemented. It topped out at 12.5Mhz.
The R800 in the Turbo R is clocked at 14.32Mhz — four times the NTSC colour burst, and like the others mentioned here, is more efficient per cycle than a Z80.
The Z380 scales up to 18Mhz, with similar considerations about Mhz not meaning the same thing as on a vanilla Z80.
The Rabbit 2000 "uses an external crystal with a frequency typically in the range from 1.8 MHz to 29.5 MHz". Which I take to mean that it tops out at 29.5 Mhz.
The eZ80 was initially announced as being intended to become available at up to 80Mhz, but it's unclear to me what actually launched. I defer to Raffzahn's existing answer which states that they only actually managed to push it to 50.