NCBI blastp seems to have a bug where it reports different top hits when -max_target_seqs is changed. This is a serious problem because the first 20 hits (for example) should be the same whether -max_target_seqs 100 or -max_target_seqs 500 is used.
The bug is reproducible on the command line when searching NCBI's nr blast database (dated 25-Nov-2015) using NCBI 2.2.28+, 2.2.30+ and 2.2.31+.
At first I thought it was something to do with my local exe/blastdb, but the same problem is also apparent on the NCBI blastp web interface (as of 30-Nov-2015)