AI Drawing Arena: GPT-5.6 Sol Beats Claude, Gemini, and Grok at Colored Pencil Art

TryAI.dev built a drawing arena that hands each model a blank white canvas and a set of colored-pencil tools—set_color, set_brush, set_pressure, draw, smudge, erase, and view_canvas—then steps back. The models either reproduce a target image (Mona Lisa, Starry Night) or draw from a text prompt. The full harness is open source at github.com/hershalb/canvas-arena.

Four vision models were tested: GPT-5.6 Sol, Claude Fable 5, Grok 4.5, and Gemini 3.6 Flash. Across two target reproductions and five open-ended prompts, 28 drawings were produced. The results are revealing.

Headline Numbers

Across all seven drawings per model, GPT-5.6 Sol averaged 29 steps and $7.74 in token cost, finishing in about 6.2 minutes. Claude Fable 5 took 41 steps, 12.5 minutes, and a staggering $160.58—over 20x the cost of Sol. Grok 4.5 used 99 steps and 34M tokens but cost only $9.21, while Gemini 3.6 Flash used 73 steps and 27.7M tokens at $12.87.

Sol never called set_color, set_brush, or set_pressure—it set those inline on each draw call, keeping its tool use lean. Grok 4.5 did the opposite: 65% of its 1,349 tool calls were those three setters. Claude Fable 5 leaned heavily on smudge (123 calls) and reviewed constantly. Gemini 3.6 Flash was the most obsessive reviewer, with nearly a third of its calls being view_canvas (about 23 self-reviews per drawing).

Quality Results

On the Mona Lisa reproduction, GPT-5.6 Sol achieved the highest SSIM score, followed by Gemini 3.6 Flash, then Claude Fable 5, with Grok 4.5 trailing. The pattern repeated for Starry Night. For the five open-ended prompts—fisherman, sunset, rose, cat, cabin—Sol's outputs were consistently more recognizable and aesthetically coherent. Claude Fable 5 produced muddy, over-smudged results. Grok 4.5's drawings were often barely legible. Gemini 3.6 Flash was solid but lacked the detail and composition of Sol.

Cost vs. Quality Tradeoff

The cost disparity is the headline. Claude Fable 5 cost $160.58 for the full run—over 20x GPT-5.6 Sol's $7.74—yet produced worse drawings. Grok 4.5 was cheap but poor. Gemini 3.6 Flash was a decent middle ground. The authors note that cheaper open-weight models were not even usable, often returning blank canvases.

This is not a benchmark of artistic ability. It's a test of how well these models handle an open-ended, long-horizon tool-use task. The results cut through the benchmaxxing and show which models can actually plan, iterate, and execute a multi-step creative process efficiently.