Emotion recognition plays a critical role in enhancing human–computer interactions. This study benchmarks the performance of commercial (GPT-4o, Claude–3.5 Sonnet) and open-source (Llama 3.1, Qwen 2.5, Mistral, Gemma 3) large language models (LLMs) on the ISEAR dataset using five prompt strategies: no persona, baseline persona, zero-shot, few-shot, and retrieval-augmented generation (RAG). Structured prompting consistently improve classification accuracy, particularly for smaller open-source models, while commercial models remained robust even under minimal prompt settings. Challenging emotions such as guilt, shame, and disgust exhibited lower classification performance across models. Model ensembling further enhanced results, with the Qwen 2.5 14B ensemble achieving the highest macro-F1 score of 78.4%. These results highlight the effectiveness of structured prompt design and model aggregation in optimizing emotion recognition. Future directions include integrating dimensional emotion frameworks and multimodal signals to build more context-sensitive and resilient systems.