The dtype of the output of np.random.choice with a scalar a is different if p is provided on windows:

>>> import numpy as np
>>> p = np.array([0.5, 0.5])
>>> np.random.choice(2, size=1).dtype
dtype('int32')
>>> np.random.choice(2, p=p, size=1).dtype
dtype('int64')
>>> np.__version__
1.13.1

I'd expect the output dtype to be consistent whether p is present or absent. Since the dtype of arange on win32 is int32, the int64 dtype output is probably the incorrect one.