What I got is just a workaround, the simple rule is Never create datetime with timezone info by using datetime().
This sample would give you a hint for this. As you see, you could avoid the unexpected difference, once and only you make "naive" datetime (it is, datetime without timezone info) and then localize it (it is not applied when you create datetime on UTC though) :
import pytz
from datetime import datetime
# make Jan 1 on PDT -> UTC
pdt = pytz.timezone("America/Los_Angeles")
pdtnow1 = datetime(2014,1,1, tzinfo=pdt)
pdtnow2 = pdt.localize(datetime(2014,1,1))
pytz.utc.normalize(pdtnow1)
# > datetime.datetime(2014, 1, 1, 7, 53, tzinfo=<UTC>)
pytz.utc.normalize(pdtnow2)
# > datetime.datetime(2014, 1, 1, 8, 0, tzinfo=<UTC>)
# make Jan 1 on UTC -> PDT
utcnow1 = datetime(2014,1,1, tzinfo=pytz.utc)
utcnow2 = pytz.utc.localize(datetime(2014,1,1))
pdt.normalize(utcnow1)
# > datetime.datetime(2013, 12, 31, 16, 0,
# > tzinfo=<DstTzInfo 'America/Los_Angeles' PST-1 day, 16:00:00 STD>)
pdt.normalize(utcnow2)
# > datetime.datetime(2013, 12, 31, 16, 0,
# > tzinfo=<DstTzInfo 'America/Los_Angeles' PST-1 day, 16:00:00 STD>)
与恶龙缠斗过久,自身亦成为恶龙;凝视深渊过久,深渊将回以凝视…