Because that's the way shells do it and File::Glob (and glob as of not-so-recent versions) uses the BSD csh (I believe) globbing implementation.

It might help to think of the character class being a predicate ("Is there a file present in the cwd which has one of these characters at the position in question?"), whereas the curly braces work more as a pattern that you're asking the globbing engine to generate irrespective of the contents of the current working directory ("Give me a list made up of items with the un-braced portion followed by each of the alternatives.").

Yes, it's nonintuitive. But that's the way it works.

Update: Durr, says right on the label: Post 5.6.0 glob is implemented with File::Glob, and I want to say the BSD-based C implementation was made at the same time. So glob should behave more-or-less the same way everywhere for not-unreasonably-old perls.

Shell example: Just as way of proof, here's the way zsh (another shell which implements csh-esque globbing) behaves. Prior to the first command there's no files in the directory.

$ zsh -c 'print foo{a,b}'
fooa foob
$ zsh -c 'print foo[a,b]'
zsh: no matches found: foo[ab]
$ touch fooa
$ zsh -c 'print foo[ab]'
fooa
[download]

Just remember that [] is a predicate on directory contents, {} is dumb pattern expansion.

Yes, glob curlies are cool and useful for things other than matching filenames. You can do stuff like this:

$pat = join '-', ('{foo,bar,baz}') x 3;
$\="\n";
print for glob $pat;
[download]

But IIRC glob() is pretty system dependent.