Contemporary foundation models match experts on text-only ophthalmology exam questions but underperform on image-based items; targeted multimodal training and prospective validation are needed before clinical deployment.