The results were published on October 25, 2026.
The CNN made the correct first-choice diagnosis in 125 of 200 cases (62%). GPT-6 Astra was correct in 84 cases and Claude Fable 5.1 in 85, both 42%. The models selected among six diagnoses from a clinical photograph and a dermoscopic image. On 30 melanomas, the CNN identified 13, versus eight for GPT-6 Astra and 11 for Claude Fable 5.1.
Lynch says the Inception v3 model, first published in 2015, was fine-tuned for about 20 minutes on 68,011 ISIC Archive images using one H100 GPU. Test lesions and other photographs of the same lesions or patients were excluded from training. The model can run in a phone browser through WebGPU without uploading images.
The source does not establish clinical safety or general diagnostic superiority. The test contains only 200 lesions, each LLM case was run once, and the author designed the evaluation and thresholds.