Old Google AI Model Beats GPT and Claude at Skin Diagnosis

Dermatologist Dr. Magnus Lynch fine-tuned Inception v3, an image-recognition architecture Google released in 2015, on 68,011 skin images using a single H100, a top-tier AI chip. In a test of 200 skin lesion cases with biopsy-confirmed diagnoses, the fine-tuned model got the right answer in 125 cases, or 62%. GPT-6 Astra, an OpenAI model, and Claude Fable 5.1, an Anthropic model, each scored around 42%, guessing 84 and 85 cases out of 200 respectively.

For one narrow task, telling types of skin lesions apart from images, a small model fine-tuned specifically for that job beat two of the most powerful general-purpose AI models available. That's an argument for picking a small, task-specific tool over a big general one, at least when the task is narrow and repeats often.