Comparison of GPT-4o and o3-Mini on Otolaryngology USMLE-Style Questions

Outcomes were to compare the accuracy of 2 large-language models-GPT-4o and o3-Mini-against medical-student performance on otolaryngology-focused, USMLE-style multiple-choice questions. With permission from AMBOSS, we extracted 146 Step 2 CK questions tagged "Otolaryngology" and stratified them by AMBOSS difficulty (levels 1-5). Each item was presented verbatim to GPT-4o and o3-Mini through their official APIs; outputs were scored correct/incorrect. Historical, de-identified student responses to the same items served as the comparator. Accuracy (%) was calculated per difficulty tier. Group differences were assessed with one-way ANOVA followed by independent-samples t tests (α=0.05). Mean accuracy across all items was 93.35% for o3-Mini and 90.45% for GPT-4o ( P =0.465). Both models outperformed students (55.44%; P =0.008 and 0.012, respectively). Performance for GPT-4o and o3-Mini remained ≥86% across all 5 difficulty levels, whereas student accuracy declined from 85.6% (level 1) to 26.7% (level 5). At the hardest tier, o3-Mini achieved 100% accuracy. GPT-4o and o3-Mini markedly exceed average medical-student performance on ENT-specific USMLE-style questions, maintaining high accuracy even at the greatest difficulty. These findings support the integration of advanced language models as adjunctive learning tools in otolaryngology.