"Drawing" the Mona Lisa with GPT-5.6, Claude, Gemini, and Grok
Summary
Analysis of TryAI's canvas-arena experiment compares four frontier models drawing Mona Lisa and Starry Night using a colored-pencil toolset. It covers tool usage, cost, token counts, SSIM-based similarity, and self-reviews, concluding GPT-5.6 Sol often outperforms others while Grok underperforms and Claude Fable is costly; the harness is open source.