Performance stability despite iteration: evaluating DeepSeek and ChatGPT on Chinese medical licensing examinations

The study compared the performance of the latest versions of large language models DeepSeek-V3.2 and ChatGPT-5.2 on the Chinese National Medical Licensing Examination (CNMLE). On a test with 600 questions from 2024, the DeepSeek series achieved 91.0% accuracy and ChatGPT 89.3%, with no statistically significant differences compared to previous versions (DeepSeek-R1 92.0% and ChatGPT-4o 87.2%). The gap between the two series narrowed from 4.8% to 1.7%. However, on the 2025 test, DeepSeek-V3.2 achieved significantly higher performance (94.7%) than ChatGPT-5.2 (89.3%). Error analysis showed no significant differences between model versions in error type classification. The results suggest that iterative updates in 2024 did not produce significant performance improvements.