In short, since I just had this discussion yesterday:

• TT is language-agnostic, and works well even in languages where whitespace matters.
• TT provides heavy customization callbacks at each level of processing
• TT has an embedded language that hides the difference between hashes and method calls so dumb hashes can be turned into smart objects later
• TT plays well outside mod_perl as a static transformation tool (or even in CGI)
• Mason has better "out of the box" caching, but TT has all the right hooks to write your own easily, probably better for real applications anyway
• TT permits Embedded Perl to be turned off or on for a given input source, useful in an enviroment where full Perl interface would be dangerous or misleading
• embedded TT triggers can be changed to suit the parsed language, and for HTML be selected so as to get into and out of WYSIWYG HTML wranglers without mangling
• One customer site I know doing $30M a year is using it and recommended it highly over the other embedded Perl/HTML solutions after doing their own study

-- Randal L. Schwartz, Perl hacker