I feel like it's more error-prone. This regexp lexer took me several minutes to get right, and I'm still not totally sure it's bug-free:

    >>> replace = lambda text, env: ''.join(env[item[1:]] if item.startswith('$') else item[1:] if item.startswith('\\') else item for item in re.findall(r'[^\\$]+|\\.|\$\w+|\\$', text))
    >>> print replace(r'This $line has \stuff \\in it that costs \$50 and some $variables.', {'line': 'LINE', 'variables': 'apples'})
    This LINE has stuff \in it that costs $50 and some apples.

If I were to write out an explicit loop over the characters of the string, I would be a lot more sure that I wasn't accidentally dropping characters due to an inadvertent failure to make the regexp exhaustive (I originally forgot the \\$ case! Although that reduces to the empty string anyway) and I wouldn't have to forget and rediscover which lexical category each token belonged to.

And, although it's not present in this case or in all regexp engines, it's a lot easier to accidentally write an exponential-time algorithm in a regexp than in a nested loop. And my experience has been that it's harder to debug it, too.