Python 3.15 base64.b64decode is slower on short inputs (wrapper overhead)
On python-build-standalone 3.15.2 vs 3.14.10 (x86-64), base64.b64decode of a 16-byte payload takes 208 ns vs 135 ns, and b64encode 100 ns vs 81 ns, while binascii.a2b_base64 itself is unchanged (47 ns). 3.15's wrapper gained padded/ignorechars/canonical and now passes four keyword arguments to a2b_base64; that alone costs 46 -> 77 ns, and the rest is sentinel/altchars handling in Python. The new C fast path only pays off above ~256 bytes (1.7-1.85x decode, 2.4-2.6x encode at >=4 KB on Sapphire Rapids, not the 3x decode claimed). If you decode many short tokens in a hot loop, call binascii.a2b_base64(s) directly.
Refined breakdown (same 3.15.2 binary, 16-byte input): 3.14's b64decode body takes 135 ns, the same body with 3.15's extra keyword-only parameters in the signature 150 ns, and 3.15's real b64decode 211 ns. So about 15 ns is the signature and about 60 ns the new body (sentinel logic plus the extra keyword arguments to a2b_base64). My '46 -> 77 ns' compared against a call with no keywords, but 3.14 already passed strict_mode=. Break-even near 256 B applies to decode only; encode is already 1.89x faster at 256 B.