A 10-year-old CNN outperforms frontier LLMs for skin cancer diagnosis

Magnus Lynch, a consultant dermatologist at Guy's Hospital, scored 200 biopsy-proven lesions from the public MILK10k dataset. Inception v3, fine-tuned on a single H100 GPU, assigned the correct diagnosis for 125 of 200 lesions, or 62 percent. GPT-6 Astra and Claude Fable 5.1 were correct for 84 and 85 lesions respectively, both reaching 42 percent. The CNN reached 173 of 200 with a second choice allowed, against 134 for GPT-6 Astra and 138 for Claude. Inception v3 dates to Google's 2015 publication, and the study used public images for fine-tuning. Lynch himself notes only 30 melanomas, single runs, possible LLM exposure to the public images, and a CNN that is a baseline rather than state of the art. Both GPT-6 Astra and Claude Fable 5.1 advised seeing a doctor for nearly every lesion when asked as a patient. For teams building AI products, the result argues for scoring the full decision, not only the label. At its preset threshold, the CNN discharged 18 of 30 melanomas. A top-line accuracy figure would hide that failure, so triage or safety-relevant features need separate tracking of misses and discharge advice.