Decade-Old CNN Beats Frontier AI Models at Skin Cancer Diagnosis

A convolutional neural network released in 2016 correctly diagnosed 125 of 200 skin lesions, beating the most powerful general-purpose AI models available as of October 2026. GPT-6 Astra got 84 right. Claude Fable 5.1 managed 85. The gap — 40 cases — sits at the heart of a new experiment in dermatologic AI.

Dr. Magnus Lynch ran the comparison. He leads the Lynch Institute, based at King's College London, where researchers work on preventing skin cancer with AI. His question was simple. Can frontier language models read a skin lesion as well as a purpose-built image recognizer?

The answer, on this evidence, is no.

How the head-to-head worked?

Three models entered. Two were general-purpose: GPT-6 Astra and Claude Fable 5.1, the strongest AI systems available at the time of testing. The third was Google's Inception v3, a convolutional neural network released in 2016 — a decade old by AI standards.

Lynch fine-tuned Inception v3 on 68,011 skin images. The process took 20 minutes. That's the entire preparation the older model received.

He then built a test set of 200 cases: 100 lesions that needed examination or treatment, and 100 benign lesions a clinician could safely leave alone and send home. Each model had to pick the most likely of six possible diagnoses.

Where the models diverged?

All three agreed on the easy cases. A basal cell carcinoma on the trunk of a 75-year-old man drew correct answers from every model.

Harder lesions split them. One invasive melanoma, under 0.8 mm thick and barely pigmented, fooled GPT-6 Astra — it answered basal cell carcinoma with 95% confidence. Claude Fable 5.1 and Inception v3 both called melanoma. A hypopigmented, thin melanoma remains exactly the lesion where a wrong confident answer matters most.

Some cases defeated everyone. A seborrheic keratosis on the leg drew "actinic keratosis" from two models and "squamous cell carcinoma" from Inception v3. All three were wrong.

What about melanoma and benign lesions specifically?

Performance varied by lesion type, and the pattern matters for triage.

• Melanoma (30 cases): Inception v3 identified 13; Claude Fable 5.1 identified 11; GPT-6 Astra identified 8.
• Nevi (50 cases): Inception v3 recognized 37; GPT-6 Astra recognized 16; Claude Fable 5.1 recognized 33.
• Seborrheic keratosis (50 cases): Inception v3 recognized 26; GPT-6 Astra recognized 13; Claude Fable 5.1 recognized 6.

The benign-lesion numbers are the striking part. Deciding what can safely be sent home is half of the test set, and it's where general-purpose models lagged furthest — a deficit with real consequences for over-referral and patient anxiety.

When each model could offer up to two candidate diagnoses instead of one, the ordering held: Inception v3 reached 173 correct answers, Claude Fable 5.1 reached 138, and GPT-6 Astra reached 134.

Does this settle the generalist-versus-specialist debate?

Lynch summarized the result plainly: "Specialized networks using older, smaller, and outdated architectures have far superior skin lesion recognition capabilities than the two state-of-the-art models."

He flagged an important caveat. Only Inception v3 received fine-tuning, and the general-purpose models ran without it. That handicap shaped the contest. He also noted that the best-performing skin cancer diagnostic AI in clinical use owes its edge to image transformation algorithms trained on large datasets — not to raw general intelligence.

The takeaway for clinicians: architecture and task-specific training still beat scale when the job is reading a lesion image. Frontier models may converse fluently about dermatology, but on this 200-case benchmark they misread one in three lesions that a 2016 network placed correctly, and the finding is likely to sharpen debate over whether general-purpose AI belongs in the skin cancer diagnostic pathway at all.