A couple of things I notice once I played with the expanded testcase. Your strings version is capturing single spaces somewhere. And both your strings and my regex are letting single $'s through as well.

I have a m/(..)/g version which avoids these two problems and is quicker than my split attempt, but still not as quick as your strings in it's current form.

our $RE_WORDS = qr[
    (?: \$? \d+ (?:[.,] \d+ )*)+
    |
    [\w\'!-]+
]xo;
sub tokenize_msg_w_m {
    my %words;
    @words{ shift =~ m[($RE_WORDS)]og } = ();
    return keys %words;
}
[download]

Results

D:\Perl\test>257026
Regex:
$ | want | 192.168.1.1 | $57 | confusing | yword |
 hopefully | combinations | LITTEL!!!!L | a | of |
 bit | is | This | to | will | text | this | 1,500 |
 Keep | out | BITH!!!!! | it | example | work | unhapp |
 $2.50 | little | MORE | I | some | thing | with | and | an

Strings:
$ |  | want | 192.168.1.1 | $57 | confusing | yword |
 hopefully | combinations | LITTEL!!!!L | a | of |
 bit | is | This | to | will | text | this | 1,500 |
 Keep | out | BITH!!!!! | it | example | work | unhapp |
 $2.50 | little | MORE | I | some | thing | with | and | an

Match:
want | 192.168.1.1 | $57 | confusing | yword |
 hopefully | combinations | LITTEL!!!!L | a | of |
 bit | is | This | to | will | text | this | 1,500 |
 Keep | out | BITH!!!!! | it | example | work | unhapp |
 $2.50 | little | MORE | I | some | thing | with | and | an

         Rate   regex   match strings
regex   526/s      --    -13%    -26%
match   607/s     15%      --    -15%
strings 710/s     35%     17%      --
[download]

The other question that crossed my mind was, what happens if the text contains "Do 33% of people watch T.V.?" ?