Ignoring the merits of repeated function calls as a benchmark, this is surprisingly good. A naive scheme implementation I wrote was about 100X slower than Python. Since Arc hasn't yet plucked the low hanging fruit of optimization, I think you folks are fine.
