First-class AI? Assessing LLMs on undergraduate physics exams

The rapid advancement of large language models (LLMs) in tackling quantitative problems presents both challenges and opportunities for science education. Here, we benchmark GPT-4.1 and GPT-5 models on a corpus of 601 undergraduate physics exam problems from the University of Bath over the past five years, marking answers against official numerical solutions. GPT-5 answered 82.9% of valid questions correctly, surpassing GPT-4.1 (65.7%) and showing roughly half the rate of question misinterpretation. Accuracy remains stable from first- to fourth-year problems and across subject areas, indicating robust performance throughout the physics undergraduate curriculum. These results position state-of-the-art LLMs at or above a first-class level, underscoring their transformative potential for assessment design in higher education.