A dermatologist compared two of the most advanced AI models available in October 2026—GPT-6 Astra and Claude Fable 5.1—against a specialized convolutional neural network (CNN) based on Inception v3, a Google model first published in 2015. The test involved 200 biopsy-proven skin lesions from the MILK10k public dataset, each photographed as both a clinical close-up and a dermoscopic (magnified) image.
The results showed a significant performance gap. The fine-tuned Inception v3 CNN correctly identified the diagnosis in 125 of 200 cases (62%), while GPT-6 Astra and Claude Fable 5.1 each correctly diagnosed 84 and 85 cases (42%) respectively. When allowing for a second-choice diagnosis, the CNN reached 173 correct answers compared to 134 and 138 for the LLMs.
The test set included 100 lesions that should not be discharged without treatment—30 melanomas, 25 basal cell carcinomas, 25 squamous cell carcinomas, and 20 actinic keratoses—plus 100 safe-to-discharge lesions of benign naevi and seborrhoeic keratoses. All lesions were biopsy-confirmed, including benign ones, making them deliberately harder to diagnose.
On melanomas specifically, the CNN identified 13 of 30 as a first-choice diagnosis, compared to 8 for GPT-6 Astra and 11 for Claude Fable 5.1. The LLMs showed systematic biases: GPT-6 Astra over-called basal cell carcinoma, while Claude Fable 5.1 favored actinic keratosis diagnoses. The CNN’s misclassifications on melanomas typically confused them with benign lesions.
The CNN used Inception v3 pretrained on everyday photographs, then fine-tuned on 68,011 dermoscopic and clinical images from the ISIC Archive. Notably, the test lesions and related patient images were removed before training. The model is small enough to run entirely in a phone browser using WebGPU, with all processing done locally.
The author notes several caveats: the test involved a relatively small number of lesions (particularly 30 melanomas), every lesion was biopsied making the cases harder than typical clinical presentations, and Inception v3 is not state-of-the-art. Contemporary specialized AI systems using ensembles of image transformers trained on larger datasets reportedly perform considerably better.
Key facts
- A fine-tuned Inception v3 CNN achieved 62% accuracy on 200 biopsy-proven skin lesions, while GPT-6 Astra and Claude Fable 5.1 each achieved 42%
- The CNN correctly identified 13 of 30 melanomas as a first-choice diagnosis versus 8 and 11 for the two frontier LLMs
- All 200 test lesions were confirmed by histopathology and included difficult cases with biopsied benign lesions
- The CNN was fine-tuned on 68,011 dermoscopic and clinical images from the ISIC Archive on a single H100 GPU
- The model runs entirely in a browser using WebGPU without uploading data to external servers