Emotion classification in low-resource languages like Bengali remains a challenging task due to data scarcity and limited pretrained resources. In this paper, we conduct a comprehensive evaluation of emotion classification models across classical machine learning, deep learning, transformer models, and instruction-tuned large language models (LLMs). Using the Bengali Emotion Corpus (BEmoC) and supplementary multilingual datasets, we assess model performance using standard classification metrics. Our results show that GPT-4 achieves the highest F1-score (82.3%), significantly outperforming the best fine-tuned transformer (XLM-R, 69.7% F1). Other LLMs such as Claude and DeepSeek also outperform traditional approaches. These findings highlight the promise of LLMs for low-resource emotion analysis while underscoring trade-offs in inference cost and deployment feasibility.