This is a bit of a nonsense benchmark—the author is basically benchmarking process startup time, and finding that if you only start one lisp process, it runs faster than if you start one jq process per input file, which runs faster than if you start one lisp process per input file. This becomes quite clear at the end when they compare multiple jq implementation (which are known to have nontrivially different parsing performance characteristics, which is the bulk of the "real work" for this task lies) and come up with execution times on the same order of magnitude.

A more apples-to-apples comparison would be to use find {} + to pass multiple filenames to jq and output using input_filename.