Generally speaking more trials per condition would render more precise (reliable or robust) estimate of the BOLD response magnitude. However, it's hard to give a magic number (the minimum number of trials) for achieving a desirable statistical power because that depends on experiment type (block or event-related), tasks/conditions, brain regions, etc..
Instead of adopting a dichotomous decision about the threshold number of trials, it would make sense to run group analysis, if this is your ultimate goal, based on the precision/reliability/robustness of each subject's beta estimate, which is what 3dMEMA was created for:
[
afni.nimh.nih.gov]
Gang