Graph bootstrap percolation, introduced by Bollobás in 1968, is a cellular automaton defined as follows. Given a "small" graph H and a "large" graph G = G 0 ⊆ K n , in consecutive steps we obtain G t+1 from G t by adding to it all new edges e such that G t ∪ e contains a new copy of H. We say that G percolates if for some t ≥ 0, we have G t = K n .For H = K r , the question about the size of the smallest percolating graphs was independently answered by Alon, Frankl and Kalai in the 1980's. Recently, Balogh, Bollobás and Morris considered graph bootstrap percolation for G = G(n, p) and studied the critical probability p c (n, K r ), for the event that the graph percolates with high probability. In this paper, using the same setup, we determine, up to a logarithmic factor, the critical probability for percolation by time t for all 1 ≤ t ≤ C log log n.The study of critical probabilities has brought numerous and often very sharp results for various graphs G and the values of the infection threshold. For example, van Enter [14] and Schonmann [20] studied r-neighbour bootstrap percolation on Z d , Holroyd [16], Balogh, Bollobás and Morris [4], Balogh, Bollobás, Duminil-Copin and Morris [3] analysed finite grids, while Balogh and Pittel [6], Janson, Łuczak, Turova and Vallier [17] and Bollobás, Gunderson, Holmgren, Janson and Przykucki [10] worked with random graphs.