There is no such thing as a cut off, because there is no such thing as not expressed - the whole genome is transcribed at some level in any given cell type. However, this doesn't stop us from sometimes needing to make a decision: which genes to include in an metagene analysis for example.
One a purely technical note, one could define a cut off as the point at which we can't distinguish between low expression levels and technical noise. It seems like zFPKM is doing something similar to this, but hand-rolled versions I've seen shift exons into nearby, but unexpressed, GC matched genome regions and then quantify them to get an average signal for strictly unexpressed sequence. I guess the difference between this and zFPKM is that zFKPM gives you a level for "unexpressed genes", where as this gives out a baseline level for "not genes".
A different approach is to think about what the level actually means. This is next to impossible with FPKM, but FPKM is readily translatable into TPM, and then the meaning is quite concrete. For example, if the average cell has 200,000 mRNA molecules in it at any one time, then a TPM of 5 would translate to 1 molecule per cell on average at any one time.
Finally, you could think in a distributional sense. I think in the last sample I looked at TPM 5 put you in the top 10% most highly expressed genes.
In the end it depends on what the purpose of the threshold is.
Thank you very much for your comments and insights ;)