This paper proposes to perform online clustering by conducting twincontrastive learning (TCL) at the instance and cluster level. Specifically, wefind that when the data is projected into a feature space with a dimensionalityof the target cluster number, the rows and columns of its feature matrixcorrespond to the instance and cluster representation, respectively. Based onthe observation, for a given dataset, the proposed TCL first constructspositive and negative pairs through data augmentations. Thereafter, in the rowand column space of the feature matrix, instance- and cluster-level contrastivelearning are respectively conducted by pulling together positive pairs whilepushing apart the negatives. To alleviate the influence of intrinsicfalse-negative pairs and rectify cluster assignments, we adopt aconfidence-based criterion to select pseudo-labels for boosting both theinstance- and cluster-level contrastive learning. As a result, the clusteringperformance is further improved. Besides the elegant idea of twin contrastivelearning, another advantage of TCL is that it could independently predict thecluster assignment for each instance, thus effortlessly fitting onlinescenarios. Extensive experiments on six widely-used image and text benchmarksdemonstrate the effectiveness of TCL. The code will be released on GitHub.