TurtleAI - Hand-Drawn: leaderboard
The hand-drawn set of TurtleAI (2026): the 102 real visual programming tasks redrawn by hand with the same solutions, testing robustness to imperfect sketches. Success means the generated program, run in a canonicalizing emulator, draws a pattern that matches the reference pixel by pixel; percent, higher is better.
Metric: Success rate (%) on the 102 TurtleAI-DSCraft tasks (hand-drawn versions of the real task images, same solutions), symbolic comparison: the predicted Python turtle program is run in a canonicalizing emulator and succeeds when its normalized drawing matches the reference pixel by pixel (0.95 or 0.92 overlap threshold), greedy decoding, one fixed prompt with the task image; higher is better. Source: arxiv.org. Saturation forecast: Around 2029. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O4 Mini (Medium) | 28.43 | #353 (O4 Mini) |
| 2 | Qwen 2.5 VL 72B Instruct | 17.65 | #846 |
| 3 | Qwen 3 VL 30B A3B Instruct | 13.73 | #871 |
| 4 | GPT-4o (2024-11-20) | 12.75 | #866 |
| 5 | GPT-4 Turbo | 8.82 | #1001 |
| 6 | O3 (Medium) | 8.82 | #228 (O3) |
| 7 | InternVL3-78B | 7.84 | #738 |
| 8 | Qwen 2 VL 72B Instruct | 7.84 | #1403 |
| 9 | Pixtral-12B | 2.94 | #2476 |
| 10 | GPT-5 (Medium) | 0.98 | #167 (GPT-5) |
| 11 | Qwen 2 VL 7B Instruct | 0 | #2553 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=turtleai-hand-drawn · How It Works · Data refreshed daily, snapshot 2026-11-13.