"Drawing" the Mona Lisa with GPT-5.6, Claude, Gemini, and Grok

I'll pull the source article so the paragraphs stay factual and specific, then draft the five–six sections you asked for.TryAI ran a colored-pencil drawing arena that handed four vision models a blank white canvas and the same toolset, then left them to draw. The models were GPT-5.6 Sol, Claude Fable 5, Grok 4.5, and Gemini 3.6 Flash. Each reproduced the Mona Lisa and Van Gogh's Starry Night under structural similarity scoring, then took five open-ended text prompts, for 28 drawings total. On Hacker News the write-up sat at 2 points with 0 comments when logged.

The harness is open source as canvas-arena. Every model got the same tools: plan as a scratchpad, view_target and view_canvas for inspection and scoring, set_color / set_brush / set_pressure for pencil state, draw for stroke batches with no solid fills, smudge for blending, and erase / clear_canvas. Reasoning was set high for every model. Target runs were scored with SSIM and RMSE after resizing both images to 256x256. Across seven drawings each, GPT-5.6 Sol averaged 29 steps and 6.2 minutes for about 3.4M tokens and an estimated $7.74; Claude Fable 5 averaged 54 steps and 12.5 minutes for 14.6M tokens and about $160.58; Grok 4.5 averaged 99 steps and 4.8 minutes for 34.0M tokens and about $9.21; Gemini 3.6 Flash averaged 73 steps and 6.9 minutes for 27.7M tokens and about $12.87.

Tool-use styles split hard. Sol never called the set_* tools and put color, brush, and pressure inline on draw, so almost all of its work was draw, smudge, and view_canvas. Grok did the opposite: about 65% of its 1,349 tool calls were set_color / set_brush / set_pressure, which is why it ran the most steps. Fable leaned on smudge and heavy review. Gemini was the most obsessive viewer, nearly a third of its calls as view_canvas, about 23 self-reviews per drawing, and like Sol it never used the set_* tools. On scored targets, every model plateaued early and finished below its best mid-run SSIM: Sol's Mona Lisa peaked at 0.352 and ended at 0.325; Gemini peaked highest overall at 0.449 on the Mona Lisa, then fell to 0.337. More reviewing did not produce a better final frame.

For engineers building long-running vision-agent loops, the useful signal is not "can a model draw" but whether self-critique and extra steps improve a closed-loop visual task. Here they often did not. Cost and latency also diverged from quality: Fable has beaten Sol on TryAI's earlier music-video arena, yet on pencils it was slower, roughly 20x more expensive than the others, and still second-tier on output. That is the product mechanic that matters for builders choosing models for multi-step tool agents: the same frontier name can invert the trade-off when the tool surface and feedback loop change.

In competitive terms, Sol was the authors' quality leader on detail for almost every drawing (exception: the elderly fisherman prompt) and their favorite outputs included its Starry Night and rose. Gemini Flash led both targets on final and peak SSIM, yet the authors note SSIM tracks structural closeness, not human preference, and said Flash did not always look closest. Grok was cheapest per token partly because roughly 98% of its tokens were cached reads, but the authors called its visual results largely unusable. Open-weight models they tried often returned a blank canvas. Fable landed second on quality at the worst cost/time trade-off of the set.

The practical takeaway is to treat open-ended tool arenas as a check on benchmaxxing, not as a ranking of artistic talent. Watch whether later Gemini frontier releases close the quality gap Flash already shows on SSIM, whether Kimi K3 changes the open-weight blank-canvas failure mode once fully open-sourced, and whether agent harnesses start stopping at peak mid-run score instead of letting models edit past their best frame. The canvas-arena repo is the concrete next step if you want the same tool contract and scoring on your own targets.