You may have missed is that it's caching a blessed scalar ref, which means that on the next new call, it can just pop an item of the cache instead of blessing a new object.

Blessing a ref takes an infinitesimal amount of time. My benchmark:

package STR;
use Class::Std::Fast; ## cache => 1;

package main;
use Time::HiRes qw[ time ];

my $start = time;
STR->new for 1 .. 1e6;
printf "Took %.8f seconds\n", ( time - $start ) / 1e6;

__END__
## With caching
c:\test>junk64
Took 0.00002447 seconds

## Without caching
c:\test>junk64
Took 0.00002391 seconds
[download]

And if you removed all the cache related conditionals, it would be quicker still.

It would also avoid your original problem