In the complex world of software systems, understanding and maintaining system stability and performance is of utmost significance. Finding anomalies in log data has become increasingly difficult due to these systems’ growing complexity. Motivated by the need to improve software release management and ensure system reliability, this study exploits Generative Pretrained Transformer (GPT)-3’s advanced word embedding and tokenizer functionalities to convert log data to adept at identifying atypical patterns and anomalies, delineated in a two-layered structure: offline and online layers. In the offline layer, historical log data undergoes processing through the GPT model, where it is divided into sentence and word embeddings. Sentence embeddings are clustered to generate labels and taggers for subsequent stages, while word embeddings directly create taggers for the online layer’s sequence labeling. The online layer involves collecting real-time data, processing it through GPT to generate embeddings, and subjecting these embeddings to a sequence labeling process. This process yields templates and variables expediting the formation of train-test data splits for a classifier that detects anomalies. Different classifiers, namely Random Forest (RF), Light Gradient Boosting Machine (LightGBM), and Categorical Boosting (CatBoost), are evaluated. Experimental analysis on four distinct real-world datasets, namely Apache, BlueGene/L (BGL), Hadoop Distributed File System (HDFS), and Thunderbird, where CatBoost achieved remarkable accuracy rates of 99.75%, 99.00%, 98.75%, and 99.33%, respectively. The study also demonstrates that GPT-based embeddings provide a more effective anomaly detection solution than Bidirectional Encoder Representations from Transformers (BERT)-based embeddings. The proposed methodology is particularly designed to be integrated into software release management processes which enables automatic anomaly detection to augment quality control measures, thereby, expediting timely intervention.