The quantum performance thing is real, but that the random circuit sampling problem they are tabling as the benchmark here is for a quantum circuit.

So really what is being claimed is that classical computers can't easily simulate quantum ones. But is that really surprising?

What would be surprising would be that kind of speedup vs classical on some kind of general optimization algorithm. I don't think that is what they are claiming though, even if it does kind of seem like it's being presented that way.