That is what we tend to do, stay close to the original voxel size. When processed with afni_proc.py, the output is (by default, at least) isotropic based on the minimum voxel dimension, truncated to 3 significant bits. Here, 3mm is already truncated, and would be used for the result.
Going from 3mm^3 to 1mm^3 is a factor of 27 in data storage, that would make a big difference. Using 3mm would not be as pretty, but it would be more honest to the data, in some sense.
- rick