My instinct is that the overhead is small. You need to add a few C stack frames and do some string conversion on each call, maybe an allocation to store the result. It’s not going to be as quick as doing in pure Rust, but the python-to-native code layer can be pretty lightweight I think!