Fake news detection is an increasingly critical challenge as misinformation proliferates across social media platforms and online news outlets. This study evaluates the effectiveness of various BERT-based transformer models—BERT base, RoBERTa, DistilBERT, ALBERT, and BERTweet—in detecting misinformation and quantifies the impact of hyperparameter optimization on their performance. Using the Optuna optimization framework, the models were fine-tuned and tested across four diverse datasets representing multiple domains, including political news, health misinformation, and general fake news, with both balanced and imbalanced distributions to simulate real-world scenarios. Results demonstrate that optimization significantly improves model accuracy and F1 scores, underscoring the importance of fine-tuning for effective fake news detection. BERT base emerged as the most optimal performer across all datasets by achieving near-perfect performance on the D4 dataset while maintaining minimal training time and resource consumption. This research contributes to the field by providing a comprehensive comparative analysis of BERT variants and highlighting the benefits of hyperparameter optimization in enhancing misinformation detection capabilities.