Oh my. This is actually so simple!
grouped = df3.groupby(level=0)
df4 = grouped.last()
df4
A B rownum
2001-01-01 00:00:00 0 0 6
2001-01-01 01:00:00 1 1 7
2001-01-01 02:00:00 2 2 8
2001-01-01 03:00:00 3 3 3
2001-01-01 04:00:00 4 4 4
2001-01-01 05:00:00 5 5 5
Follow up edit 2013-10-29
In the case where I have a fairly complex MultiIndex, I think I prefer the groupby approach. Here's simple example for posterity:
import numpy as np
import pandas
# fake index
idx = pandas.MultiIndex.from_tuples([('a', letter) for letter in list('abcde')])
# random data + naming the index levels
df1 = pandas.DataFrame(np.random.normal(size=(5,2)), index=idx, columns=['colA', 'colB'])
df1.index.names = ['iA', 'iB']
# artificially append some duplicate data
df1 = df1.append(df1.select(lambda idx: idx[1] in ['c', 'e']))
df1
# colA colB
#iA iB
#a a -1.297535 0.691787
# b -1.688411 0.404430
# c 0.275806 -0.078871
# d -0.509815 -0.220326
# e -0.066680 0.607233
# c 0.275806 -0.078871 # <--- dup 1
# e -0.066680 0.607233 # <--- dup 2
and here's the important part
# group the data, using df1.index.names tells pandas to look at the entire index
groups = df1.groupby(level=df1.index.names)
groups.last() # or .first()
# colA colB
#iA iB
#a a -1.297535 0.691787
# b -1.688411 0.404430
# c 0.275806 -0.078871
# d -0.509815 -0.220326
# e -0.066680 0.607233