I would like to propose (purely for convenience reasons) to bump the minimal numpy version to 1.7.0. There are a couple of hacks left to support numpy 1.6 (category.py,tests/test_category.py,cbook.py (putmask vs copyto), possibly(?)colors.py), and 1.7 also introduced the ability to pass a tuple as theaxis kwarg to reducing ufuncs.

numpy 1.7.0 was released in Feb. 2013 (3y 9m ago). As a comparison, the minimal version of numpy was bumped to 1.6.0 in Jul. 2014 (#3229), 3y 2m after its release (May 2011) (and was effective with the release of mpl1.4.0 in Aug. 2014).

Edit: I misread the release notes of 1.7.0 -- turns out that what it added was the ability to passmultiple axes as a tuple to reducing ufuncs; passing a single axis was supported since long ago (and that was one of my bigger reasons). So the issue is less pressing on my side at least.