[ScottBurson]

There's no free lunch.

While GC requires tuning for the best performance, (pure) reference counting imposes a different cost: because it can't collect circular structures, the programmer has to break cycles herself. This clutters the application code. So you can't simply take a program written to run on a GC platform and move it to a reference counting platform unmodified.

Alternatively, you could use both: do most of your reclamation with reference counting, but throw in an occasional GC to collect circular structures. CPython, the most commonly used Python implementation, does this. I doubt it is the best thing for performance, but CPython isn't very fast anyway.