Third-generation sequencing, also called long-read sequencing, is revolutionizing genome assembly: as PacBio and Nanopore technologies become more accessible in technicity and in cost, long-read assemblers flourish and are starting to deliver chromosome-level assemblies. However, these long reads are also error-prone, making the generation of a haploid reference out of a diploid genome a difficult enterprise. Although failure to properly collapse haplotypes results in fragmented and/or structurally incorrect assemblies and wreaks havoc on orthology inference pipelines, this serious issue is rarely acknowledged and dealt with in genomic projects, and an independent, comparative benchmark of the capacity of assemblers and post-processing tools to properly collapse or purge haplotypes is still lacking. To fill this gap, we tested different assembly strategies on the genome of the rotifer Adineta vaga, a non-model organism for which high coverages of both PacBio and Nanopore reads were available. The assemblers we tested (Canu, Flye, NextDenovo, Ra, Raven, Shasta and wtdbg2) exhibited strikingly different behaviors when dealing with highly heterozygous regions, resulting in variable amounts of uncollapsed haplotypes. Filtering out shorter reads generally improved haploid assemblies, and we also benchmarked three post-processing tools aimed at detecting and purging uncollapsed haplotypes in long-read assemblies: HaploMerger2, purge_haplotigs and purge_dups. Testing these strategies separately and in combination revealed several approaches able to generate haploid assemblies with genome sizes, coverage distributions, and completeness close to expectations.