I've used cd-hit and think that does the job reasonably well.

A study found that the best strategy is first the novo assembly with Trinity, followed by CAP3 to eliminate redundancy:

Optimizing de novo assembly of short-read RNA-seq data for phylogenomics BMC Genomics 2013, 14:328 http://www.ncbi.nlm.nih.gov/pubmed/23672450

only use these programs if you know what you are doing. I usually use CD-HIT-EST straight after with a 95% similarity level to remove highly similar sequences then I follow up with cap3. You should read the literature and get a feel for the settings used and try to udnerstand why they use such settings. You can also filter by FPKM thresholds, but have a good reason for it. I think Trinity has accompanying scripts for that. Usually the lower supported reads are the ones that are fine to remove, but depending on what your study aim is, you may not want to remove such reads with low counts.