Performance of 5 AI Models on United States Medical Licensing Examination Step 1 Questions: Comparative Observational Study

TL;DR

This study compared five AI models on USMLE Step 1 questions, finding Grok had the highest accuracy and consistency, especially with images and complex questions.

Contribution

The study provides a comparative evaluation of newer AI models on USMLE Step 1 questions, highlighting their strengths in different medical domains and question formats.

Findings

Grok achieved the highest score (91.6%) and excelled in image-based and case-based questions.

DeepSeek scored lowest initially but matched others on text-only questions.

Models performed best in biostatistics and worst in musculoskeletal topics.

Abstract

Artificial intelligence (AI) models are increasingly being used in medical education. Although models like ChatGPT have previously demonstrated strong performance on United States Medical Licensing Examination (USMLE)–style questions, newer AI tools with enhanced capabilities are now available, necessitating comparative evaluations of their accuracy and reliability across different medical domains and question formats.

This study aimed to evaluate and compare the performance of 5 publicly available AI models: Grok, ChatGPT-4, Copilot, Gemini, and DeepSeek, on the USMLE Step 1 free 120-question set, assessing their accuracy and consistency across question types and medical subjects.

This cross-sectional observational study was conducted between February 10 and March 5, 2025. Each of the 119 USMLE-style questions (excluding 1 audio-based item) was presented to each AI model by using a standardized prompt cycle. Models answered each question 3 times to assess confidence and consistency. Questions were categorized as text-based or image-based and as case-based or information-based. Statistical analysis was performed using chi-square and Fisher exact tests, with Bonferroni adjustment for pairwise comparisons.

Grok achieved the highest score (109/119, 91.6%), followed by Copilot (101/119, 84.9%), Gemini (100/119, 84%), ChatGPT-4 (95/119, 79.8%), and DeepSeek (86/119, 72.3%). DeepSeek’s lower score was due to an inability to process visual media, resulting in 0% accuracy on image-based items. When limited to text-only questions (n=96), DeepSeek’s accuracy increased to 89.6% (86/96), matching Copilot. Grok showed the highest accuracy on image-based (21/23, 91.3%) and case-based questions (70/78, 89.7%), with statistically significant differences observed between Grok and DeepSeek on case-based items (P=.01). The models performed best in biostatistics and epidemiology (5.8/6, 96.7%) and worst in musculoskeletal, skin, and connective tissue (4.4/7, 62.9%). Grok maintained 100% consistency in responses, while Copilot demonstrated the most self-correction (112/119, 94.1% consistency), improving its accuracy to 89.9% (107/119) on the third attempt.

AI models showed varying strengths across domains, with Grok demonstrating the highest accuracy and consistency in this dataset, particularly for image-based and reasoning-heavy questions. Although ChatGPT-4 remains widely used, newer models like Grok and Copilot also performed competitively. Continuous evaluation is essential as AI tools rapidly evolve.

Figures