A "drawing arena" was created where AI models, including GPT-5.6 Sol, Claude Fable 5, Grok 4.5, and Gemini 3.6 Flash, were tasked with reproducing images like the Mona Lisa and creating original art from text prompts. The models used a set of colored-pencil tools, making decisions on color, tip width, pressure, and when to smudge or erase, while periodically viewing their canvas to assess progress.

Claude Fable 5 proved to be the most expensive, costing an estimated $160 for its drawings, significantly more than the others. Despite this, and beating GPT-5.6 in previous challenges, it performed worse on this drawing task. Models showed a tendency to plateau early in their drawing process, often exceeding their best performance before finishing. For instance, Gemini 3.6 Flash achieved the highest peak similarity score but ended with a lower score.

The study highlights that while cheaper, open-weight models can handle execution tasks, frontier models excel at more complex, open-ended challenges. The arena's open-source nature allows others to replicate these experiments. The findings also emphasize that more frequent self-reviewing did not necessarily lead to a better final output.