"Drawing" the Mona Lisa with GPT-5.6, Claude, Gemini, and Grok
2 hours ago
- Four frontier AI models (GPT-5.6 Sol, Claude Fable 5, Grok 4.5, Gemini 3.6 Flash) were tasked with drawing the Mona Lisa and other images using a colored-pencil toolset on a blank canvas.
- GPT-5.6 Sol was the overall best performer, producing the most detailed and appealing drawings, while Grok 4.5 performed poorly with mostly unusable output.
- Claude Fable 5 was the most expensive and slowest model, costing roughly 20x more than others yet not matching GPT-5.6's quality.
- Gemini 3.6 Flash scored highest on structural similarity (SSIM) for target reproductions but showed large score fluctuations and frequent editing.
- The experiment highlights the gap between frontier and open-weight models, reveals tool usage patterns, and demonstrates that more reviews do not guarantee better final results.