CCBench: leaderboard
About 180 feature tasks in private CodeCrafters codebases (e.g. add authentication to a Redis clone, backreferences to a regex engine), graded by CodeCrafters' test runners; success rate (2026).
Metric: Accuracy (%). Source: huggingface.co. Status: saturated. 229 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | InternVL3-78B | 83.5 |
| 2 | InternVL3-38B | 80.6 |
| 3 | InternVL3-8B | 77.8 |
| 4 | InternVL2-8B | 77.1 |
| 5 | GPT-4.1 (2025-04-14) | 73.5 |
| 6 | Qwen 2 VL 72B | 69.8 |
| 7 | Qwen 2 VL 7B | 65.7 |
| 8 | GPT-4.1 Mini | 64.3 |
| 9 | Grok 2 (1212) | 57.3 |
| 10 | Claude 3.5 Sonnet (20241022) | 56.1 |
| 11 | Claude 3.7 Sonnet | 55.1 |
| 12 | Claude 3.5 Sonnet | 54.1 |
| 13 | Llama 3.2 90B Vision Instruct | 54.1 |
| 14 | Qwen 2 VL 2B | 53.7 |
| 15 | Gemma 3 27B | 53.3 |
Interactive version: theaggregate.ai/benchmark?slug=ccbench · How It Works · Data refreshed daily, snapshot 2026-09-05.