T The rapid growth of social media, particularly x (twitter), produces large-scale public opinion data that are informal and highly contextual, posing challenges for sentiment analysis. this study evaluates the performance of convolutional neural network (cnn), bidirectional long short-term memory (bilstm), and two hybrid architectures (cnn-bilstm and bilstm-cnn) for sentiment classification of tweets discussing the free nutritious meal (mbg) policy, using indobert embeddings as frozen feature representations. a quantitative comparative experimental design with stratified 10-fold cross-validation was employed on 1,569 indonesian tweets after case folding, cleansing, and normalization. models were assessed using accuracy, precision, recall, f1-score, roc-auc, computational efficiency (training and inference time), and t-sne visualization to examine latent feature separability. the results indicate that hybrid models achieve more stable and competitive performance than single architectures. cnn excels in computational efficiency, while bilstm and hybrid models better capture sequential context. t-sne plots reveal clearer class separation for hybrid architectures. it is concluded that, under a fixed indobert embedding, hybrid architectures especially cnn-bilstm and bilstm-cnn provide the best trade-off between predictive performance and efficiency for indonesian x (twitter) sentiment analysis.