You 'orig' variation never sets $t, but I don't think that matters much. Anyway, I made a slight variation, to reduce the chance the PRNG favours one method over the other:

use strict;
use warnings;
     
use Benchmark qw( cmpthese );
        
for our $n ( map 10**$_, 1, 3, 6 ) {
    print "Benchmarking $n values\n";
    my $l = int log $n;
    our @data = ();
    our @target = ();
    foreach (1 .. $l) {
        push @data, [map int rand $n, 1 .. $n];
        push @target, int rand $n;
    }
    cmpthese -log $n, {
        orig => q[
            for (my $i = 0; $i < @data; $i++) {
                my $c = 0;
                for my $elem (@{$data[$i]}) {
                    $c++ if $elem > $target[$i];
                }
                my $pct= (@{$data[$i]} - $c ) / @{$data[$i]} * 100;
            }
        ],
        sort => q[
            for (my $i = 0; $i < @data; $i++) {
                my $c = 0;
                my @n = sort {$a <=> $b} @{$data[$i]};
                $c++ while $n[$c] <= $target[$i];
                my $pct= (@{$data[$i]} - $c ) / @{$data[$i]} * 100;
            }
        ],
        grep => q[
            for (my $i = 0; $i < @data; $i++) {
                my $pct = grep({$_ > $target[$i]} @{$data[$i]})
                          / @{$data[$i]} * 100;
            }
        ],
    };
}
[download]

This shows 'sort' to be always slower, and 'grep' always winning:

Benchmarking 10 values 
        Rate sort orig grep
sort 34448/s   -- -11% -53%
orig 38503/s  12%   -- -47%
grep 72551/s 111%  88%   --
Benchmarking 1000 values
      Rate sort orig grep
sort 162/s   -- -57% -62% 
orig 381/s 135%   -- -11%
grep 427/s 164%  12%   --
Benchmarking 1000000 values
            (warning: too few iterations for a reliable count)
            (warning: too few iterations for a reliable count)
            (warning: too few iterations for a reliable count)
     s/iter sort orig grep
sort   31.2   -- -83% -85%
orig   5.18 502%   -- -10%
grep   4.67 567%  11%   --
[download]