We consider the problem of testing bipartiteness in the adjacency matrix model. The best known algorithm, due to Alon and Krivelevich, distinguishes between bipartite graphs and graphs that are ∊-far from bipartite using \mathop 0\limits^ \sim (1/\varepsilon ^2 ) queries. We show that this is optimal for non-adaptive algorithms, up to polylogarithmic factors. We also show a lower bound of \Omega (1/\varepsilon ^{3/2} ) for adaptive algorithms.