In today’s advanced communication systems, including 5G and beyond, the use of automatic modulation recognition or classification (AMC) has emerged as an increasing need. In recent years, research on deep learning-based AMC methods, which are preferred over traditional methods, has gained significant importance. In this study, the aim is to investigate the effects of optimization algorithms, also known as optimizers, used in deep learning on AMC. For this purpose, three fundamental architectures are considered in the study: CNN, LSTM, and GRU. The seven widely recognized optimizers, namely SGD, Adagrad, RMSprop, Adadelta, Adam, Adamax, and Nadam, are utilized during the training process of each network, based on the underlying architectures. By training the networks on the RML2016.10a dataset using these optimizers, a total of 21 distinct deep learning models are generated. The models are assessed based on key performance indicators, including training time, accuracy, and inference time. It has been observed that the performance of the optimizers varies depending on the networks and key performance indicators. The results are presented in a comparative manner.