Hacker News front pageJul 22, 2026, 05:13 UTC+8

Four frontier models draw the Mona Lisa with colored pencils

"Drawing" the Mona Lisa with GPT-5.6, Claude, Gemini, and Grok

TryAI built a canvas arena where GPT-5.6 Sol, Claude Fable 5, Grok 4.5, and Gemini 3.6 Flash used colored-pencil tools to reproduce the Mona Lisa and Starry Night, plus five open-ended prompts. GPT-5.6 Sol scored highest on SSIM; Claude Fable 5 took the longest and cost the most while producing worse output. Grok 4.5 struggled, and open-weight models returned blank canvases. The authors argue fuzzy tasks like this separate frontier models from the rest better than benchmarks, and reveal real costs of long-running agent work.

Why it matters: TryAI built a drawing arena where GPT-5.6 Sol, Claude Fable 5, Grok 4.5, and Gemini 3.6 Flash used colored-pencil tools to reproduce famous paintings — 28 drawings total, with cost and structural similarity scores. It's a rare hands-on test of tool use + visual feedback loops,...

Read the original ↗Export Markdown

Take

TryAI built a colored-pencil canvas and had four models reproduce the Mona Lisa and Starry Night, plus five open-ended prompts — 28 drawings total. Models could pick colors, tip width, pressure, smudge, erase, and call view_canvas to check their work. GPT-5.6 Sol scored highest on SSIM; Claude Fable 5 took the longest, cost the most, and produced worse output. Grok 4.5 struggled, and several open-weight models returned blank canvases.

The reason to click: this turns "what makes frontier models better" into something you can see. Not a benchmark number — actual drawings. Claude Fable 5 beat GPT-5.6 on their earlier music-video challenge but flopped here, which tells you model rankings flip depending on the task shape. If you're picking a model for long-running agent work, "slower, pricier, and worse" is a trade-off worth knowing.

I'd discount this a bit: SSIM measures structural similarity, not artistic quality; the open-ended prompts had no objective scoring, just eyeballing. The post doesn't give exact dollar amounts or per-run durations — only that Fable 5 was "much more expensive and slower." If you run this yourself, your cost and timing may differ a lot.

Source: Hacker News front page · tryai.dev
Published: Jul 22, 2026, 05:13 UTC+8 · Jul 22, 2026