"Drawing" the Mona Lisa with GPT-5.6, Claude, Gemini, and Grok
Source Entity
Hacker News
An experimental study tested the artistic capabilities of four frontier AI models by tasking them with creating physical-style drawings. The research evaluated performance through objective scoring of reproductions and open-ended creative prompts.
The AI Artistic Frontier: Benchmarking Generative Vision
A recent experimental study has pushed the boundaries of artificial intelligence by evaluating the creative and motor-planning capabilities of four leading frontier models: GPT-5.6 Sol, Claude Fable 5, Grok 4.5, and Gemini 3.6 Flash. By placing these models in a controlled 'drawing arena,' researchers sought to move beyond text-based generation and into the realm of iterative, tool-assisted physical representation. The process involved providing the models with a blank canvas and a digital suite of colored-pencil tools, allowing them to manage color, stroke width, pressure, and blending.
The Mechanics of Machine Creativity
Unlike standard image generation where an AI outputs a complete pixel grid in one pass, this experiment forced the models to engage in a recursive loop. The models were required to lay down strokes, smudge for blending, and utilize an 'erase' function, punctuated by a 'view_canvas' call to inspect their work. This feedback loop mimics the human artistic process of trial and error, testing whether these models can maintain a coherent vision while managing the long-term state of a project.
Methodology and Scope
To ensure objective assessment, the researchers selected two iconic targets: Leonardo da Vinci’s Mona Lisa and Vincent van Gogh’s Starry Night. By scoring these reproductions against the originals, the team could quantify the spatial awareness and color fidelity of each model. Furthermore, the inclusion of five open-ended prompts allowed the researchers to assess how these models handle abstract creative tasks when they are not anchored to a specific ground-truth reference image.
Economic and Efficiency Analysis
Beyond the aesthetic quality of the drawings, the study tracked the 'dollar cost' of each session. In the current landscape of Large Language Models (LLMs), the cost of inference is a critical metric for developers. By tracking every stroke and tool call, the researchers provided a granular look at the economic viability of using AI for complex, multi-step artistic workflows that require high-frequency model interaction.
Evaluating Improvisation
One of the most intriguing aspects of the study is the focus on whether these models 'actually improvise.' Improvisation requires the model to deviate from a plan when it detects a mistake or discovers a more effective technique during the drawing process. This distinguishes true cognitive-like planning from simple rote execution, offering a window into how these models handle the unpredictability of creative output.
Future Implications for Generative AI
This research represents a significant shift from static image synthesis to autonomous agent-based creation. As models move toward more sophisticated tool use, the ability to plan, execute, and self-correct will become a hallmark of advanced artificial intelligence. The findings from this experiment will likely influence how future versions of these models are trained to handle complex, multi-modal tasks that require physical-world reasoning and iterative refinement.