This also happen in 2.0.0rc2

http://matplotlib.org/2.0.0rc2/_images/hexbin_demo.png

However this is not a regression, it has been like this for a long time: http://matplotlib.org/1.2.1/_images/hexbin_demo.png

This is the ever-present problem of filling adjacent patches--the same as the problems with contourf. If the edges are not stroked at all, there are light artifacts. If the edges are stroked, as has always been the choice in hexbin, then the patches overlap, which is what is we see here. Turning anti-aliasing on or off modifies the artifacts, and their behavior also changes depending on alpha, but the upshot is that we have never found a solution for this problem. It seems to be a general problem with renderers. For hexbin, using a thin edgewidth helps.

https://lists.cairographics.org/archives/cairo/2007-July/011098.html may be relevant (to explain the issue, not to solve it).

It may be possible to fix this issue by drawing the hexbin as a quadmesh rather than a polycollection.

I don't think so. First, hexagons are not quadrilaterals. Second, the quadmesh is not immune to the adjacent patch rendering problems.

1. A hexagon can be represented as two quadrilaterals (in fact, splitting each hexagon in a top and a bottom half actually gives you a proper quadmesh).
2. I see no seams in https://files.gitter.im/matplotlib/matplotlib/9dTT/default-quadmesh.png (produced by Agg), so I assume that Agg (and cairo) do have special handling for them. (At least on cairo's side, the entire quadmesh is defined before rendering, so it should have all the info to compute the seams between the quads.)

1. Yes, of course one can do this.
2. The Agg rendering seems to work well with and without antialiasing, provided alpha is 1. It does not do so well with alpha = 0.3, for example. Still, an improvement for the alpha=1 case is worthwhile, so I like your idea.
   

It seems there is an inaccuracy in agg quadmesh, though:

x = np.arange(4)
y = np.arange(3)
z = np.arange(6).reshape(2, 3)
X, Y = np.meshgrid(x, y)
pcolormesh(X+Y, X-Y, z)
savefig('skewed.png')

All of the corners are slightly chopped off, with the top corner being the most obvious.

Closing as duplicate of #1188.