Try asking GPT-5.6, Claude, Gemini, and Grok to “draw” the Mona Lisa using code (SVG/Canvas) and see which one best captures the proportions of the face, hands, and lighting compared to the original. Then find out which AI is best suited for this kind of task.
It sounds like a strange test, but it actually measures a model’s reasoning quite well. Drawing an image through code requires converting a “mental image” purely into numerical coordinates—there’s no direct image generation like with Midjourney.
I think this is exactly where the differences become clear. Some models understand what kind of smile the Mona Lisa has, what angle she’s viewed from, but when converted into code, it falls apart and doesn’t look like a person at all.
This article will walk through the actual results from each AI, along with which one is best suited for this kind of creative-coding work.
Mona Lisa Drawing Results from Each AI
When you see the four actual images lined up together, the differences are much clearer than expected. Some capture the facial structure, proportions, and lighting close to the original, while others end up as geometric shapes instead of a person.
GPT-5.6 and Claude seem to understand the “layout” of the image better than the others, positioning the hands, face, and background in a balanced way. Gemini and Grok, on the other hand, still struggle with distorted facial proportions.
I think this really reflects the models’ spatial reasoning capabilities, since they have to convert a mental image into coordinates and render it out while still keeping it looking like a person—not just generating a pretty image the way a diffusion model does.
The Origin of This Strange Experiment
The idea started with a simple question: what would happen if you asked an AI chatbot to draw an image without using an image generation model?
Instead of asking Midjourney or DALL-E to create a Mona Lisa image, we tried having GPT-5.6, Claude, Gemini, and Grok write SVG code or ASCII art instead. This is a test of how well an AI truly understands an “image” in its “mind”—not just recreating patterns memorized from training data.
Honestly, I think this is a much harsher test than typical image generation, because language models aren’t trained specifically to draw. They have to convert their understanding of shapes, proportions, light, and shadow into code that, once rendered, still looks like the Mona Lisa.
The results were more interesting than expected.
Where Each AI Stands in the Current LLM War
The LLM market is fiercely competitive right now, with each company touting its own selling points.
OpenAI’s GPT-5.6 emphasizes deeper reasoning, thinking step-by-step before answering. Claude from Anthropic is known for its accuracy in coding tasks and following complex instructions.
Google’s Gemini claims to be strong in multimodal capabilities, having been trained to understand images, text, and code simultaneously from the start, rather than stitching them together afterward. Grok from xAI focuses on speed and real-time updated information.
I think each one claims different strengths, but a test like drawing the Mona Lisa reveals weaknesses that the marketing never mentions, because it measures genuine “understanding”—not just the ability to answer questions correctly.
When AI Has to Interpret Art as Code
This skill of converting images into code isn’t just for show—it has plenty of real-world uses. For example, creating diagrams to quickly explain workflows to a team, or generating data visualizations from just a description, without needing to open Excel.
Another clear use case is helping design rough UI drafts—type out where you want a button, what color, and the AI can assemble it into code right away.
The most interesting case is using it to teach children how to draw. Have the AI convert a description into an image, and the child sees how their words turn into shapes—training both language skills and spatial thinking at the same time.
I think this is exactly what sets it apart from previous generations of AI, because it’s not just about “recognizing” the Mona Lisa—it has to interpret proportions, light, and shadow into code that actually runs. These capabilities can be applied immediately to design or education work, no waiting required.
Who Draws the Mona Lisa More “Convincingly”?
| Factor | GPT-5.6 | Claude |
|---|---|---|
| Facial Proportions | Close to the original | Slightly too elongated |
| Lighting and Shadow | Smooth gradation | Rather flat |
| Composition | On point | On point |
| Factor | Gemini | Grok |
|---|---|---|
| Creativity | Minimalist interpretation | Added background detail |
| Facial Proportions | Somewhat distorted | Close to the original |
| Understanding of Light and Shadow | Basic | Basic |
Honestly, judging from the rendered code output, GPT-5.6 understood “lines” and lighting the best in this group, while Grok excelled at adding atmosphere that made the image feel more alive.
I don’t think any of them drew it 100% like the original, but just seeing an AI interpret a world-famous work of art into code to this extent is already impressive enough.
Pros and Cons of Each Model for This Kind of Task
When you actually compare them, each model has its own distinct strengths and weaknesses—none of them excels across the board.
GPT-5.6 draws lines and composes images most accurately, but takes longer to process than the others. Gemini, on the other hand, is fast with concise code, but still misses some facial details.
Pros
- +GPT-5.6 interprets lighting, shadow, and facial proportions closest to the original
- +Grok adds background atmosphere with more dimension, giving the image a lively feel
Cons
- −Claude doesn't render small details like hands or hair as clearly as it should
- −Gemini works fast, but the overall image still lacks fine linework
I think if accuracy is the priority, GPT-5.6 is the way to go. But if you want fast results for quickly testing ideas, Gemini fits the bill better.