We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
Hacker News·14d ago·5 min read
"Drawing" the Mona Lisa with GPT-5.6, Claude, Gemini, and Grok
TL;DR
A drawing arena was created to evaluate four AI models—GPT-5.6, Claude, Grok, and Gemini—by having them reproduce famous artworks and respond to open-ended prompts. Results showed significant differences in performance, highlighting the capabilities and limitations of each model.
✦ Why It Matters
Experiment with different AI models to find the best fit for creative tasks in your projects.
Key Takeaways
- Use GPT-5.6 for efficient and high-quality drawing tasks.
- Evaluate cost versus output quality when selecting AI models.
- Monitor model performance to avoid diminishing returns in iterations.
- Explore open-source tools for custom drawing tasks.
Full Summary
Researchers built a drawing arena where AI models could create art using a set of colored-pencil tools. They tested four models—GPT-5.6 Sol, Claude Fable 5, Grok 4.5, and Gemini 3.6 Flash—on two famous artworks, the Mona Lisa and Van Gogh's Starry Night, as well as five open-ended prompts.
Each model's performance was evaluated based on structural similarity scores (SSIM) and the time and cost of execution. GPT-5.6 Sol scored highest in quality and efficiency, with a cost of $7.74 for seven drawings, while Claude Fable 5 was the most expensive at $160.58 but produced lower quality results.
Notably, models tended to plateau in quality early in the drawing process, indicating that more iterations did not necessarily lead to better outcomes. These findings highlight the varying capabilities and costs of different AI models in creative tasks.