Hi @avtober,
It is very hard to make a general recommendation that works on every dataset and so in that paper we did our best to make a recommendation that should be ok in most situations. Don’t feel bound by that recommendation and your best bet is to fit what works best for your data.
However, I wanted to add a further recommendation. Even though Deblur can operate on pre-merged reads, it probably is not your best option here. This is because as read-length increase, the number of sequences you retain with Deblur drops dramatically, see here for an example of the numbers you might see with 300 bp, and then consider your case that has much longer reads. Deblur was really designed to work with with shorter, single end reads, it performs really well in those cases. However, if you are working with paired-end reads, I would recommend going with DADA2, which does not penalize read lengths to the same extend. So you would end up with loads more reads. There are lots of posts on this forum about selecting truncating parameters with DADA2 and merging paired-end reads in mind.