Dermatologist Tested GPT-6 Astra and Claude on 200 Skin Lesions: A 2015 CNN Was More Accurate

⚡ The gist in 5 seconds - Dermatologist Magnus Lynch compared GPT-6 Astra, Claude Fable 5.1, and a fine-tuned Inception v3 on 200 skin lesions with biopsy-confirmed diagnoses. - The decade-old convolutional network Inception v3 from 2015 won on accuracy. - The test breakdown was published on Habr as a translation; the specific metrics and methodology are in the original article. ### 🔍 What was found Magnus Lynch, a practicing dermatologist, collected 200 skin lesions, each with a biopsy-confirmed diagnosis. He ran this dataset through three systems: GPT-6 Astra, Claude Fable 5.1, and a fine-tuned Inception v3 convolutional network from 2015. The result: a decade-old narrowly specialized CNN outperformed both cutting-edge multimodal models. The test design is the classic "narrow task vs. generalist" setup. Inception v3 was fine-tuned specifically for dermatological images, whereas GPT-6 Astra and Claude Fable 5.1 operate as general-purpose models. The full per-class figures for each type of lesion and the error analysis are in the Habr translation, linked in the source. ### 💡 Why it matters The "large multimodal model vs. small specialized one" debate plays out precisely where mistakes are costly — in image-based diagnostics. Lynch's test shows that a fine-tuned CNN on a narrow domain still holds its own, and a general multimodality benchmark is no substitute for validation on a specific clinical task. Whether the same picture would hold on other datasets — the source makes no claim; this is a single sample limited to 200 cases.