With KEEPALIVE_SIZE_LIMIT = 200, the pybench runtime is basically the
same as with my patched version. stringbench remains a bit faster though
(~8%).

You say that RAM size is cheaper than CPU power today, which is true but
misses one part of the picture: the importance of CPU caches, and thus
of working set size. There are also probably people wanting to use
Python in memory-constrained environments (embedded).

I understand the argument about possible optimizations with an external
buffer, but are such optimizations likely to be implemented? (see
#1590352 and #1629305). If they really are, then I'm happy with the
unicode type remaining a plain PyObject!