This must be OS dependent or something... (hard to say as you didnt post the output). On my system print with no cache is much faster:

open my $zap, ">", "NULL:";
use Benchmark "cmpthese";
cmpthese -10, {
   every => sub { for (1..100) { print $zap $_ } },
   tens  => sub { my $x;
                  for (1..100) {
                    $x .= $_;
                    unless ($x % 10) {
                      print $zap $x;
                      $x="";
                    }
                  }
                  print $zap $x if length $x;
                },
}

__END__
# modfied code as posted above (5.6.1 on W2k)
         Rate  tens every
tens   1159/s    --  -90%
every 11285/s  874%    --

# original code (5.6.1 on W2k)
         Rate  tens every
tens   1124/s    --  -90%
every 11219/s  899%    --

#  5.8.4 (XP)
         Rate  tens every
tens   8305/s    --  -73%
every 30329/s  265%    --

# 5.8.6 No Implicit Sys (XP)
         Rate  tens every
tens   9655/s    --  -68%
every 29755/s  208%    --
[download]

At least on Win32 it would seem just using print is much faster.... I guess this could be an example of what Steve Hay was talking about with Win32's realloc being crap and that it makes concatenation unecessarily slow. Update: No, that doesnt make sense, I just tried it with two different perl versions on two different win32 boxes and "every" won every one of them....