I think it's some sort of interesting how small the gap is.

GIL isn't really relevant for blocking IO-bound code, because the lock is released (and not contended) when blocking. For nonblocking single-threaded code such as Tornado's event loop, the GIL also doesn't matter. Not to mention that the real speed concerns had nothing to do with interpreted versus compiled, because Tornado very clearly held its own, and web.py+flup (etc.) have architectural issues with regard to performance, far beyond mere interpreter overhead.