$^$^$^_ and $^$^$^z are both recognized by zxcvbn as bruteforce regions. it reports the entropy as:

n log (c)

for a length-n password with symbol space c. the huge difference in crack time is because zxcvbn is using c==33 (symbols only) for $^$^$^_ and c==59 (symbols + a-z) for $^$^$^z

$^$^$^i is in the middle -- 'i' is considered a dictionary match, the rest is c==59 bruteforce.

the bigger problem is $^$^$^ isn't recognized as a pattern, but i'm working on ways to improve bruteforce estimation too. good example!

One of the cases I tried was:

notaword12 - 17 hours

notawrod123 - 22 minutes

With almost any not-word, really. The extra 12 is better than 123 somehow.