Inception v3 Tops Frontier LLMs on Skin Lesions

The real lesson is less “old beats new” than “specialization beats model prestige” in tightly scoped medical imaging tasks.

• –The CNN achieved 62% accuracy versus 42% for both frontier multimodal LLMs across six diagnoses.
• –It identified 13 of 30 melanomas, compared with 8 for GPT-6 Astra and 11 for Claude Fable 5.1.
• –The 200-case, author-reported test is too small for clinical conclusions, and each LLM was run only once.
• –Safety thresholds matter: the CNN’s default setting discharged more melanomas, despite stronger diagnostic accuracy.
• –Running locally in a phone browser suggests a compelling privacy and latency advantage for specialized medical vision models.