your post slightly conflates two different issues: what is the difference and which annotation is better

finding differences is not so hard, gffcompare for example or simple alignments can do it

finding out, overall which assembly is better is far more difficult as both builds might have strengths and weaknesses and those tradeoffs need to be scored in some way

scientists have been producing genome builds over extended periods, years - a new, better build takes a long time to hone

One way to assess the quality of annotations is to match the genome builds to sequencing reads, transcript sequences or protein sequences deposited in various repositories.