On October 25, 2026, Dr Magnus Lynch, a consultant dermatologist and Mohs surgeon at Guy's Hospital and honorary reader at King's College London, published a comparison on his lab's website. He designed the test and made the clinical rulings. He also discloses that "The analysis and the first drafts were produced with Claude Code, Anthropic's coding tool, which runs on a Claude model", and that one of the two tested models is Anthropic's. This is a single-author web report, not a peer-reviewed study.
The setup
The test set is 200 randomly drawn lesions from MILK10k, a public International Skin Imaging Collaboration dataset released in 2025. Every lesion is confirmed by histopathology and has a clinical photo and a dermoscopic image. Half should not be discharged: 30 melanomas, 25 basal cell carcinomas, 25 squamous cell carcinomas and 20 actinic keratoses. Half are safe to discharge: 50 naevi and 50 seborrhoeic keratoses. GPT-6 Astra and Claude Fable 5.1 ran at default settings, saw both images and nothing else, and chose from six diagnoses or "unsure", with a confidence, a second choice and a discharge decision. The comparison network is Inception v3, published by Google in 2015, which Lynch "fine-tuned all layers on 68,011 dermoscopic and clinical images from the ISIC Archive on a single H100 GPU", with the test lesions removed from training.
The results
On first-choice diagnosis, the CNN was right on 125 of 200 lesions (62%), against 84 for GPT-6 Astra and 85 for Claude Fable 5.1 (42% each). On first or second choice it scored 173, against 134 and 138. For melanoma specifically, first-choice hits were 13, 8 and 11 of 30. Lynch's summary: "A ten-year-old network, fine-tuned for about twenty minutes, beat both of them at recognising skin cancer."
The discharge decision tells a different story. The two LLMs kept 97 and 96 of the 100 unsafe lesions, discharging 3 and 4 of the 30 melanomas, but they discharged only 32 and 22 of the 100 safe lesions. At its preset threshold the CNN discharged 76 safe lesions but kept only 77 unsafe ones, and it would have discharged 18 of 30 melanomas. At a second, "matched" threshold, which the article does not define further, it kept 98 or 99 unsafe lesions and discharged 2 or 1 melanomas; no safe-lesion figure is given for that setting. In a separate patient-style test using only the clinical photo, the LLMs advised seeing a doctor "for 99 and 98 of a further 100 biopsy-proven melanomas."
What remains uncertain. Lynch lists the limits himself. Two hundred lesions and 30 melanomas leave wide uncertainty, and each LLM case was run once. All lesions were biopsied, so the set is harder than routine images: the CNN's area under the ROC curve was 0.96 on ordinary archive images against 0.86 here. The LLMs may have seen MILK10k in training, which would favour them. And he calls Inception v3 a baseline; leading systems use transformer ensembles trained on more data.
Analysis: The headline is about recognition, and there the result is clear: a small model trained on the task beat general models that were never tuned for it. But a triage tool is judged at an operating point, and the numbers show the two kinds of system failing in opposite directions. The LLMs behaved like cautious generalists, safe on cancers and of little use for reducing referrals. The CNN discriminated better, yet at its preset threshold it would have sent most melanomas home. For anyone evaluating models for a screening workflow, the lesson is to report the full trade-off: accuracy, misses on the dangerous class and how many harmless cases are cleared, at a threshold fixed before testing. If a threshold is tuned after seeing test results, the outcome is a comparison figure, not a deployment figure. The broader engineering point holds across domains: before reaching for a frontier model on a narrow perception task, measure a fine-tuned specialist, because it may be cheaper, faster and better.
Sources and evidence
An October 25 report by dermatologist Magnus Lynch found a 2015 Inception v3, fine-tuned for about twenty minutes on 68,011 ISIC images, named 125 of 200 biopsy-proven lesions correctly against 84 for GPT-6 Astra and 85 for Claude Fable 5.1. But at its preset threshold the CNN would have discharged 18 of 30 melanomas, while the LLMs discharged 3 and 4 and cleared only 32 and 22 of 100 harmless lesions. Judge triage models at a fixed operating point, not on accuracy alone.