CodeEditorBench Plus - Code Debug (Three-shot): leaderboard
Code Debug task of CodeEditorBench (Guo et al., 2024) on its Plus split, the problems kept after timestamp-based filtering against training-data contamination: the model repairs a program seeded with one to four errors, and passes when the fixed program passes every test case on the authors' online judge. Scored as pass@1 (%) with greedy decoding and three in-context examples; higher is better.
Metric: pass@1 (%, greedy decoding, three-shot, CodeEditorBench_Plus). Source: codeeditorbench.github.io. Saturation forecast: Estimated already saturated. 19 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4 (0613) | 34.5 | #1214 |
| 2 | GPT-3.5 Turbo (1106) | 27 | #3245 |
| 3 | Magicoder-S-DS-6.7B | 26.2 | #7023 |
| 4 | Phind-CodeLlama-34B-v2 | 23.9 | #9928 |
| 5 | GLM-4 | 23.3 | #1673 |
| 6 | Gemini 1.0 Pro | 22.9 | #4110 |
| 7 | CodeLlama-7B-Instruct-hf | 16.7 | #9444 |
| 8 | CodeLlama-13B-Instruct-hf | 16 | #9291 |
| 9 | Magicoder-S-CL-7B | 15.7 | #9042 |
| 10 | CodeLlama-34B-Instruct-hf | 14.3 | #7428 |
| 11 | CodeLlama-34B-hf | 13.3 | #9332 |
| 12 | WizardCoder-15B-V1.0 | 11.4 | #9717 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=codeeditorbench-plus-code-debug-three-shot · How It Works · Data refreshed daily, snapshot 2026-11-12.