Abstractive summarization plays a critical role in managing large volumes of textual data, yet it faces persistent challenges in consistency and evaluation. Our study compares two state-of-the-art models, PEGASUS and Flan-T5, across a diverse range of benchmark datasets using both ROUGE and BARTScore metrics. Findings reveal that PEGASUS excels in generating detailed, coherent summaries for large-scale texts evidenced by an R-1 score of 0.5874 on Gigaword while Flan-T5, enhanced by our novel T5 Dual Summary Framework, produces concise outputs that closely align with reference lengths. Although ROUGE effectively measures lexical overlap, its moderate correlation with BARTScore indicates that it may overlook deeper semantic quality. This underscores the need for hybrid evaluation approaches that integrate semantic analysis with human judgment to more accurately capture summary meaning. By introducing a robust benchmark and the pioneering T5 Dual Framework, our research advocates for task-specific optimization and more comprehensive evaluation methods. Moreover, current dataset limitations point to the necessity for broader, more inclusive training sets in future studies.