Johnny Chadda

Mona Lisa

The same painting, prompt and effort, compared across AI models.

On this page

I thought it would be fun to make AI models paint something familiar to everyone, yet challenging enough so there is room to grow for the future. This is where I gather their versions of the Mona Lisa for comparison.

The benchmark

Every model gets the same prompt through Opper, with reasoning_effort: "high":

Paint the Mona Lisa as an SVG. Use viewBox=“0 0 800 1200”.

Each painting is the model’s first result, without edits or retries. The SVG is rendered as an image for display here. New models get a short post with their own painting, and their result is added to this page.

The captions show the recorded time, output tokens and cost for each run. Gemini 4 Carbon was run on November 11, 2026; the other models were run on November 10.

Paintings

The Mona Lisa as painted by Gemini 4 Carbon
Gemini 4 Carbon · 258 s · 21,230 output tokens · $0.21
The Mona Lisa as painted by Claude Fable 5.2
Claude Fable 5.2 · 397 s · 35,607 output tokens · $1.42
The Mona Lisa as painted by Claude Opus 5.5
Claude Opus 5.5 · 162 s · 15,866 output tokens · $0.32
The Mona Lisa as painted by GPT-6 Astra
GPT-6 Astra · 363 s · 13,775 output tokens · $0.69
The Mona Lisa as painted by GPT-6.1 Sol
GPT-6.1 Sol · 201 s · 9,357 output tokens · $0.094
The Mona Lisa as painted by Gemini 4 Argon
Gemini 4 Argon · 359 s · 30,065 output tokens · $0.30
The Mona Lisa as painted by Kimi K3.1 Pro
Kimi K3.1 Pro · 280 s · 9,589 output tokens · $0.096
The Mona Lisa as painted by DeepSeek V4.1 Flash
DeepSeek V4.1 Flash · 47 s · 9,579 output tokens · $0.012
The Mona Lisa as painted by GLM 5.3
GLM 5.3 · 717 s · 43,621 output tokens · $0.19

I think the results are quite interesting, and while most models get the high level right, there is plenty of room for improvement.

Related