Drawing the Mona Lisa with GPT-5.6 Sol, Claude Fable 5, Grok 4.5, Gemini 3.6 Flash – Results and Analysis

Why we run these drawing arenas

We run these open‑ended drawing tasks because they are fun, visually informative, expose the frontier‑vs‑open gap, show the real cost of long‑running tasks, and generate interesting community discussion.

The tools available to every model

All four models received the same colored‑pencil toolset: plan (scratchpad), view_target, view_canvas, set_color, set_brush, set_pressure, draw (batch of strokes), smudge (blend region), erase, and clear_canvas. The harness is open source at github.com/hershalb/canvas-arena.

How the models used the tools

GPT-5.6 Sol never called set_color, set_brush, or set_pressure; it set those values inline on each draw call, so its calls were mostly draw, smudge, and view_canvas. Grok 4.5 did the opposite: 65% of its 1,349 tool calls were set_color / set_brush / set_pressure, which explains its high step count. Claude Fable 5 relied heavily on smudge (123 calls) and reviewed the canvas 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) and, like Sol, never touched the set_* tools.

Headline numbers across all seven drawings per model

• GPT-5.6 Sol: 7 drawings, 29 steps, 6.2 min, 3.4 M tokens, $7.74 cost, 6.0 average metric, 116 peak metric.
• Claude Fable 5: 7 drawings, 54 steps, 12.5 min, 14.6 M tokens, $160.58 cost, 15.6 average metric, 207 peak metric.
• Grok 4.5: 7 drawings, 99 steps, 4.8 min, 34.0 M tokens, $9.21 cost, 4.3 average metric, 354 peak metric.
• Gemini 3.6 Flash: 7 drawings, 73 steps, 6.9 min, 27.7 M tokens, $12.87 cost, 22.6 average metric, 190 peak metric.

Objective similarity for the Mona Lisa target

Model	Steps	Time	Reviews	Final SSIM	Peak SSIM	RMSE
GPT-5.6 Sol	44	8m41s	7	0.325	0.352	54.2
Claude Fable 5	86	18m58s	27	0.286	0.289	56.9
Grok 4.5	106	4m30s	5	0.151	0.152	95.6
Gemini 3.6 Flash	60	5m31s	22	0.337	0.449	51.1

Objective similarity for the Starry Night target

Model	Steps	Time	Reviews	Final SSIM	Peak SSIM	RMSE
GPT-5.6 Sol	23	5m47s	4	0.130	0.179	37.8
Claude Fable 5	43	11m30s	12	0.153	0.176	40.9
Grok 4.5	58	5m51s	3	0.039	0.039	104.9
Gemini 3.6 Flash	74	6m36s	24	0.172	0.190	42.8

Did the models actually improve during a run?

Two patterns held across both targets: first, models plateaued early—Claude Fable 5’s Mona Lisa similarity was essentially flat after about the fifth review despite 27 total reviews. Second, in every target run the final drawing scored below the best score reached mid‑run. For example, GPT-5.6 Sol’s Mona Lisa peaked at 0.352 SSIM and ended at 0.325; its Starry Night peaked at 0.179 and ended at 0.130. Gemini 3.6 Flash reviewed the most (about 23 times per drawing) and reached the highest peak of the batch, 0.449 on the Mona Lisa, then fell to 0.337 by the end. More reviewing did not translate into a better finish; the models kept editing past their own best performance.

Cost and token usage

Claude Fable 5 cost an estimated $160 for its seven drawings, roughly 20 × the cost of GPT-5.6 Sol ($7.74), Grok 4.5 ($9.21), or Gemini 3.6 Flash ($12.87). Grok used the most tokens (34 M) but stayed cheap because ~98 % were cached reads at a fraction of the input rate, and Grok’s rates are the lowest of the four. Gemini also leaned heavily on cached reads, keeping its cost modest despite modest despite 27.7.7M token count.

Our take

GPT-5.6 Sol was the runaway leader on quality and efficiency; its Starry Night and rose drawings were our favorites. Grok 4.5 produced mostly unusable output despite frantic tool use. Gemini 3.6 Flash was better than Grok but still trailed Sol in final similarity while incurring higher cost. Claude Fable 5 placed second on quality but at a far higher cost and time, making it the least appealing trade‑off for this task.

Synthesis of community comments

Commenters noted the visual character of the outputs. One observed that the drawings looked "childish," with some models rendering concepts like shading and refraction while others used flat colors for glass, stem, and rose. Another remarked that the Grok results were "amusing, almost comically bad" and likened them to early Microsoft Paint attempts. A few users found the Grok images "uncanny and surreal," evoking strong negative feelings, while one compared them to "a truly disturbed child" like the drawings from The Ring. Several praised GPT-5.6 Sol’s efficiency, highlighting the stark contrast in token usage and cost versus Claude Fable 5 (3.4 M vs 14.6 M tokens, $7.74 vs $161). Some questioned the choice of SSIM/RMSE as a feedback metric, arguing that it may degrade artistic output and suggesting perceptual metrics like DINOv2 cosine similarity could yield better results. A few commenters suggested adding other models such as Kimi, Deepseek, or GLM to future arenas, and one noted that the models appeared to lack a "revert to previous" capability, causing them to keep adding strokes to fix earlier mistakes rather than stepping back.