RNA-Seq de novo assembly is an important method to generate transcriptomes for non-model organisms before any downstream analysis. Given many great de novo assembly methods developed by now, one critical issue is that there is no consensus on the evaluation of de novo assembly methods yet. Therefore, to set up a benchmark for evaluating the quality of de novo assemblies is very critical. Addressing this challenge will help us deepen the insights on the properties of different de novo assemblers and their evaluation methods, and provide hints on choosing the best assembly sets as transcriptomes of non-model organisms for the further functional analysis. In this article, we generate a textquotedblleftreal timetextquotedblright transcriptome using PacBio long reads as a benchmark for evaluating five de novo assemblers and two model-based de novo assembly evaluation methods. By comparing the de novo assmblies generated by RNA-Seq short reads with the textquotedblleftreal timetextquotedblright transcriptome from the same biological sample, we find that Trinity is best at the completeness by generating more assemblies than the alternative assemblers, but less continuous and having more misassemblies; Oases is best at the continuity and specificity, but less complete; The performance of SOAPdenovo-Trans, Trans-AByss and IDBA-Tran are in between of five assemblers. For evaluation methods, DETONATE leverages multiple aspects of the assembly set and ranks the assembly set with an average performance as the best, meanwhile the contig score can serve as a good metric to select assemblies with high completeness, specificity, continuity but not sensitive to misassemblies; TransRate contig score is useful for removing misassemblies, yet often the assemblies in the optimal set is too few to be used as a transcriptome.