Benchmarking is always tricky: providing subrefs to benchmark makes it easier to get the code right, but if the target code is very fast the sub call overhead can swamp the results. Providing strings instead gets rid of that overhead, but is far harder to get right because the code will be evalled in a different context.

Using a string version of the benchmark, I confirm your results on my system perl (5.28) but the difference almost entirely disappears with 5.34 (which is also quite a bit faster overall):

% perl benchmark
               Rate methodic   blessy   direct
methodic 16017006/s       --     -31%     -40%
blessy   23189661/s      45%       --     -13%
direct   26673243/s      67%      15%       --

% /opt/v5.34.0/bin/perl benchmark
               Rate methodic   blessy   direct
methodic 21275455/s       --     -39%     -39%
blessy   34927866/s      64%       --      -0%
direct   35023414/s      65%       0%       --

% cat benchmark
use strict;
use warnings;
use Benchmark;

our($mcount, $bcount, $dcount) = (0) x 3;
package Methodic {
    sub new { return bless {} }
    sub method { ++$::mcount }
};
package Blessy {
    sub new { return bless sub { ++$::bcount } }
};
sub direct { ++$::dcount }

our $methodic = Methodic->new;
our $blessy = Blessy->new;
our $direct = \&direct;

Benchmark::cmpthese(-1, {
    methodic => q{$::methodic->method()},
    blessy => q{$::blessy->()},
    direct => q{$::direct->()},
});
%
[download]

Note also that if you look at the counts after the benchmark has run, you'll see larger number than those that were reported. IIRC this is because Benchmark tries to calculate the overhead of calling and adjust results for it, and should not be taken as a sign that it can't count. :)