Rapid Evolution of Large Language Models in Medical Education: Comparative Performance of ChatGPT-3.5, ChatGPT-5, and DeepSeek on Medical Microbiology MCQs

Rapid advances in large language models (LLMs) warrant specialty-specific benchmarking to assess their educational potential and limitations. We evaluated the newly released generative artificial intelligence (genAI) model ChatGPT-5, DeepSeek-R1, and the early ChatGPT-3.5 on 80 multiple-choice questions (MCQs) from a medical microbiology course examination, weighted for midterm and final components. Items were classified according to the revised Bloom’s taxonomy. Performance was compared with that of more than 150 Doctor of Dental Surgery students. Content quality was assessed independently by two consultants in clinical microbiology using the validated CLEAR tool modified to assess AI content completeness, accuracy, and relevance. The mean total scores were 80.5 for ChatGPT-3.5, 96.0 for ChatGPT-5, and 95.5 for DeepSeek, versus a student mean of 86.21/100. ChatGPT-5 and DeepSeek-R1 significantly outperformed ChatGPT-3.5 in completeness and accuracy scores, with no differences between them. ChatGPT-5 maintained high accuracy across lower- and higher-order cognitive Bloom’s domains, whereas DeepSeek-R1 showed a significant drop in higher-order items. For ChatGPT-3.5, incorrect responses had longer answer-choice word counts. CLEAR scores were significantly higher for correct versus incorrect responses in all models (p < 0.001). This study showed that the currently available LLMs can exceed average student performance in medical microbiology while providing high-quality explanations. Regular benchmarking is essential to ensure responsible integration of genAI into educational, pedagogical, and assessment tools.