It's looking as if Open Babel is taking the second to last step from Gaussian geometry optimization output.
For instance, a CCSD(T)/QZVP optimization on the Li_2^+ ion gives a final bond length of 3.133200 Å when you look at the log file, but extracting the geometry from the log file with OpenBabel 2.3.2 I get 3.13236 Å.
Log file attached.