CodeARC - Anonymized: leaderboard
Reverse-engineering a hidden Python function: the agent sees input and output examples, may query the function on inputs of its choosing and call a differential-testing oracle, and must write code that passes a differential test against it within the query budget. 1,114 functions with their names anonymized so they give nothing away. Higher is better. Source: arxiv.org/abs/2503.23145.
Metric: Success Rate (%; anonymized function names, 1,114 functions). Source: arxiv.org. Saturation forecast: Estimated already saturated. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O3 Mini | 52.7 | #283 |
| 2 | O1 Mini | 47.7 | #335 |
| 3 | DeepSeek R1 | 41.3 | #266 |
| 4 | Claude 3.7 Sonnet | 33.8 | #263 |
| 5 | DeepSeek V3 | 29.5 | #334 |
| 6 | GPT-4o | 28.7 | #358 |
| 7 | Llama 3.1 405B Instruct | 26.7 | #472 |
| 8 | Qwen 2.5 72B Instruct | 21.6 | #470 |
| 9 | Llama 3.1 70B Instruct | 20 | #570 |
| 10 | Llama 3.2 90B Vision Instruct | 19.7 | #626 |
| 11 | GPT-4o Mini | 18.5 | #601 |
| 12 | Llama 3.2 11B Instruct | 16.1 | #1110 |
| 13 | Qwen 2.5 7B Instruct | 15.8 | #866 |
| 14 | QwQ-32B | 15.4 | #422 |
| 15 | Mixtral 8x22B Instruct | 15 | #741 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #6 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=codearc-anonymized · How It Works · Data refreshed daily, snapshot 2026-11-07.