We study an online saddle point problem where at each iteration a pair of actions need to be chosen without knowledge of the future (convex-concave) payoff functions. The objective is to minimize the gap between the cumulative payoffs and the saddle point value of the aggregate payoff