The answer is 5, except when it's not. Most arguments for using more test participants are wrong, but some tests should be bigger and some smaller.
