Anomaly detection is a field of research necessary in various practical scenarios such as cyber fraud detection and medical diagnosis. There have been many anomaly detection studies using deep self-supervised learning, but relatively few investigations on tabular data tasks. In our study, we propose Contrastive Learning with Anomaly Generation (CLAG), an effective learning method that utilizes generated anomaly. This is a combination of anomaly generation scheme with contrastive representation learning, which has been in the spotlight recently. In our experiments, CLAG shows high performance on multiple tabular datasets, especially state-of-the-art performance on Thyroid. We analyze which factors are helpful in learning the representations required for anomaly detection through several additional experiments.