This is a regression in 2.0b3 compared to 1.5.2 (haven't bisected more)

from matplotlib import pyplot as plt
import numpy as np

np.random.seed(1)
fig, (ax1, ax2) = plt.subplots(2)
for x, y in np.random.rand(10, 2):
    ax1.text(x, y, "foo", clip_on=True)
    ax2.text(x, y, "foo")
plt.show()

1.5.2 displays everything fine

2.0b3 clips away all text objects in the top axes

(so I'm marking this as 2.0 but feel free to change it)

The texts actually appear to be clipped to the "wrong" place and use the wrong zorder: you can see them appear "behind the axes" if you pan the axes as below: