For older versions of GNUbg (upto v0.90, mid 2013), 3-ply checkerplay was better than 2-ply, but 3-ply cube action was actually quite a bit worse than 2-ply. Overall, the 3-ply cube action was bad enough that 3-ply could be playing worse than 2-ply overall.
The newer versions of GNUbg (v0.91 and v1.0 and up) play better in general and have solved much of this "bad odd-ply cube action" problem. Now 3-ply cube action is just a little bit better than 2-ply cube action and 3-ply checker play is still clearly better, so overall 3-ply is clearly better than 2-ply.
Some figures from a benchmark, showing the errors made:
Bot Description Chequer Errors Cube Errors All Errors
--------------------------------------------------------------------------------------------------------------------------------------------------------------
GNUBG v0.91 4-ply 9.1071(1138) 3.4287( 116) 12.5358(1254)
GNUBG v0.91 3-ply Grandmaster 12.8651(1391) 3.9101( 134) 16.7752(1525)
GNUBG v0.91 2-ply WorldClass 16.6188(1559) 4.3317( 135) 20.9505(1694)
GNUBG v0.90 4-ply 16.0467(1414) 8.8001( 195) 24.8468(1609)
GNUBG v0.90 2-ply WorldClass 25.9465(1814) 10.3792( 212) 36.3257(2026)
GNUBG v0.90 3-ply Grandmaster 22.0227(1794) 17.8564( 269) 39.8791(2063)
--------------------------------------------------------------------------------------------------------------------------------------------------------------
Bad 3-ply cube action, and resulting higher error total of the older v0.90 version in red
So, for the old gnubg, 2-ply was a better setting than 3-ply, or one could use 3-ply checker with 2-ply cube to get the best of both worlds.
The new gnubg is simple: 4-ply is better than 3-ply is better than 2-ply, for cube, checker and overall.