Note that this effect varies across builds and versions. For most of Python's history, we could give clear and invariant optimization advice (locals and nonlocals are fastest, global variable access was at least twice as slow, and builtin variable access was even slower). That ordering will likely remain true but absolute speeds have improved dramatically and the ratios have shifted).
Here is a run of Tools/scripts/var_access_benchmark.py for Python 3.12rc1 stock build for an Apple M1 Max (your mileage may vary):
Variable and attribute read access:
1.9 ns read_local
2.4 ns read_nonlocal
2.8 ns read_global
4.1 ns read_builtin
5.0 ns read_classvar_from_class
12.1 ns read_classvar_from_instance
4.8 ns read_instancevar
4.7 ns read_instancevar_slots
12.2 ns read_namedtuple
29.0 ns read_boundmethod
Variable and attribute write access:
2.4 ns write_local
2.5 ns write_nonlocal
10.5 ns write_global
26.8 ns write_classvar
4.3 ns write_instancevar
4.2 ns write_instancevar_slots
Data structure read access:
5.7 ns read_list
11.1 ns read_deque
10.1 ns read_dict
10.5 ns read_strdict
Data structure write access:
6.2 ns write_list
11.3 ns write_deque
11.2 ns write_dict
12.0 ns write_strdict
Stack (or queue) operations:
18.4 ns list_append_pop
17.8 ns deque_append_pop
18.1 ns deque_append_popleft
