Classifier gradient penalty GAN is a GAN proposed to perform self-supervised class-conditional data generation and clustering on unlabeled datasets. The classifier gradient penalty GAN's generator takes a continuous latent vector and a categorical latent vector as input and generates a class-conditional data point corresponding to the categorical latent vector. In this paper, we propose to leverage the codebook architecture to improve the performance of classifier gradient penalty GAN. In the proposed architecture, the generator takes the page vector of the codebook corresponding to the index of the categorical latent vector, instead of taking the one-hot categorical latent vector directly. Unlike the codebook used in generative models with vector quantization, the codebook of the proposed architecture is not embedded with the encoder. Instead, the codebook is simply trainable and updated via generator loss like trainable parameters in the generator. The proposed architecture improved the quality of the generated data, class-conditional data generation performance, and clustering performance of the classifier gradient penalty GAN.