The online learnability of binary and multiclass hypothesis classes is very well understood. In the noise-free setting, it is well known that only two minimax error rates are possible: constant or scaling linearly with the time horizon T. In this talk, we show that slight modifications to the standard online classification game can result in very different minimax rates. In particular, we present two natural modifications to the standard online learning setup which result in a trichotomy of minimax rates. That is, in these settings, the minimax error rates scale only like one of three different functions of the time horizon T.