It requires the data to be normally distributed.

It constructs the null distribution by reshuffling group labels.

It provides an exact p-value for the observed statistic.

It is only valid for large sample sizes $n > 30$.