This study compares the performance of ChatGPT and DeepSeek in mental health education for university students. Using a culturally adapted Mental Health Question Prompt Manual, responses to 21 prompts were collected across 30 iterative rounds. Word embeddings and Frobenius similarity were employed to evaluate the semantic consistency and diversity of the responses. The results indicate that ChatGPT produces more stable outputs, while DeepSeek demonstrates greater variability and adaptability. The findings highlight the importance of optimized prompt design in enhancing the educational value of generative AI in mental health contexts.