We introduce DrVD-Bench, the first benchmark to evaluate vision-language models' clinical reasoning ability across graded medical imaging tasks.