This very much depends on which benchmark you pick. I did quite an extensive benchmark comparison between Flecs and EnTT, and they outperform each other in different areas: https://github.com/SanderMertens/ecs_benchmark

I am not trying to start an argument over which model is better. In fact, I am a big proponent of hybrid models- and I think I have been pretty upfront about the disadvantages of archetypes in this post.