Deep learning (DL) has potential in the diagnosis of novel coronavirus disease (COVID-19). Nevertheless, the effectiveness of DL models varies when applied to different datasets and cross-datasets. This paper proposes optimized neural architecture search network (NASNet) for COVID-19 diagnosis. Two forms of NASNet models, namely NASNet-Mobile and NASNet-Large are applied to a dataset of 3411 computed tomography (CT) lung images freely available on GitHub repository. For the experimentation, 85% of the total samples are used for training, while the remaining are used for testing. The training and testing losses and the classification accuracies are varied with respect to the number of epochs. Results show that at an epoch of 15, NASNet-Mobile has an accuracy, a recall and an area under the receiver operating characteristics curve (AUC) of 82.42%, 78.16% and 91.00%, respectively. On the other hand, NASNet-Large has an accuracy, a recall and an AUC of 81.06%, 80.43% and 89.00%, respectively.