Four AIs drew the Mona Lisa. All of them edited it worse.

TryAI built a drawing arena: a blank canvas and a colored-pencil toolset with every shortcut stripped out. No fills, no shapes — just pick a color, set tip width and pressure, lay strokes, smudge, erase, and call view_canvas to check your work. GPT-5.6 Sol, Claude Fable 5, Grok 4.5, and Gemini 3.6 Flash each took a run at the Mona Lisa and Starry Night.

The technique story is real but less mystical than the telling: with only strokes and smudging available, all four converged on layering and blending to build tone — the way anyone actually using pencils has to. The tools forced the method, not some latent art-history instinct.

The genuinely interesting result is the self-editing. Gemini peaked at 0.449 SSIM mid-run and finished at 0.337. Every model plateaued early and ended below its own best frame. Gemini reviewed its canvas ~23 times per drawing; Grok looked ~4 times — and more looking bought nothing. Cost spread was brutal too: GPT-5.6 Sol finished top of the pile for $7.74, while Claude Fable 5 burned $160.58 for a worse score.

Why it matters: agents that can see their own output still don't know when to stop — and "let it iterate" is a default worth questioning.