It would seem that loop bookkeeping dominates execution time here. Without it, I get quite different results:

    >>> timeit.timeit("'%s' % i", setup="i = 42", number=1000000)
    0.31799793243408203
    >>> timeit.timeit("str(i)", setup="i = 42", number=1000000)
    0.4146881103515625

This is another argument for profiling everything.