STAR is better in most ways, from mapping accuracy to speed. The big caveat to STAR is that you need a good bit of RAM. For a nice objective look at STAR and other RNAseq aligners, I would recommend that you have a quick read through this recent and very thorough comparison from the RNA-seq Genome Annotation Assessment Project in Nature Methods (there's a similar comparison by the same collaboration for transcript reconstruction in the same issue).

BTW, the take-home message from that paper can probably be summed up from Figure 3 (the paper is open access, so this is a direct link)

Edit: Have a look at IV's answer as well. I hadn't mentioned Gsnap, but I can also say that it's always produced very good results if you have an annotation (this seems to be confirmed in the review that I linked to).

To my opinion, some of the most important pros and cons:

Tophat

Pros

• Widely used + huge community to ask questions in fora
• No fuss connections with cufflinks and any other Tuxedo pipeline tool
• A great part of published results are based on this aligner and is widely accepted
• Provides a ready to use junction file

Cons

• Really slow response rates from the relevant helpdesk email
• Doesn't do read clipping for partial read alignment (which is really useful in many scenarios)
• Inner mate distance and sd have to be calculated beforehand for optimal performance

Star

Pros

• Super fast
• The latest versions get really good statistics in comparisons
• Can do read clipping
• It has a mode of output compatible with cufflinks
• Provides a ready to use junction file

Cons

• The first versions had many issues.
• Not so many users as other aligners but I think that there is a strong user base, especially after ENCODE
• I'll add in the list also GSNAP, which we also widely use in the lab

GSNAP

Pros

• Always one of the best in any comparison article out there (usually 1st or 2nd)
• Can handle partial matches with clipping and indels
• Not that resource intensive
• The creator really supports the aligner and answers really fast in emails
• Gets splicing junctions really well
• Has a mode compatible with cufflinks
• Provides multiple sam output (concordant, halfmapping, paired, halfmapping, unique, multimapping, etc)

Cons

• The last version (with the suffix array) is a lot faster than any previous version but still slower than Star [unless it's run within the ultrafast algorithm max allowed mismatches]
• Not so many users as Tophat (even though you can get also really good feedback from Trinity users)

STAR is much faster than Tophat

TopHat is more widely used, and if you need help with it, there are a lot more users who can help. (see how many people use the TopHat tag over the STAR tag)

STAR is much faster than Tophat, but I don't know which result is more reliable. But I think both are OK