There's no correct answer. Also, you should not be using Cufflinks - it has been updated to StringTie'. Moreover, FPKM is not an ideal expression unit to use in terms of comparisons across samples. If you are analysing TCGA data, you neither have to use the FPKM counts, in most cases, as the raw count HTseq files are available for download.
With that, my swift answer is to eliminate genes whose mean FPKM value are below 10. However, due to the fact that there is no cross-sample normalisation performed when deriving FPKM counts, a value of 10 means different things in different samples.