Currently, the axis offset text is always typeset with the format string "%1.10e" followed by some crazy postprocessing inScalarFormatter._formatSciNotation. I would suggest switching to some "%g"-based format, so that e.g. 1234 doesn't get formatted as "1.234e3" which can hardly be described as "more legible". Probably_formatSciNotation could be simplified a bit too.

Simple example:

python -c 'from pylab import *; plot(*[1234 + arange(10)] * 2); show()'

Edit: the above is now fixed by the switch of axes.formatter.offset_threshold to 4, but the issue remains with

python -c 'from pylab import *; plot(*[12.321 + arange(5) * .001] * 2); show()'

where the offset text should be "12.320" (or at least "12.32"), not "1.232e1".