Code editing benchmark: 225 Exercism exercises across 6 languages (C++, Go, Java, JS, Python, Rust). Measures ability to edit existing code, not just generate new code.
Tests real-world code editing across multiple languages; uses Exercism exercises with test suites; measures practical coding ability beyond generation.
Results depend on edit format (whole file vs diff); cost varies dramatically by model; Aider-specific prompting may favor some models.
Higher is better. Percentage of exercises where the model's edits pass all test cases (pass@2). Top models score 50-70%.
Max score: 100
Results depend on edit format (whole file vs diff).
| # | Model | Score |
|---|---|---|
| 1 | 94.0 | |
| 2 | 92.2 | |
| 3 | 91.9 | |
| 4 | 91.6 | |
| 5 | 90.8 | |
| 6 | 90.6 | |
| 7 | 90.3 | |
| 8 | 90.2 | |
| 9 | 89.1 | |
| 10 | 89.0 | |
| 11 | 88.6 | |
| 12 | 88.5 | |
| 13 | 88.5 | |
| 14 | 88.5 | |
| 15 | 88.2 | |
| 16 | 88.1 | |
| 17 | 88.0 | |
| 18 | 87.3 | |
| 19 | 87.1 | |
| 20 | 87.0 | |
| 21 | 86.7 | |
| 22 | 86.5 | |
| 23 | 86.5 | |
| 24 | 86.3 | |
| 25 | 86.0 | |
| 26 | 85.8 | |
| 27 | 85.7 | |
| 28 | 85.7 | |
| 29 | 85.6 | |
| 30 | 85.6 | |
| 31 | 85.4 | |
| 32 | 85.3 | |
| 33 | 85.0 | |
| 34 | 84.9 | |
| 35 | 84.8 | |
| 36 | 84.8 | |
| 37 | 84.6 | |
| 38 | 84.4 | |
| 39 | 84.2 | |
| 40 | 84.1 | |
| 41 | 84.0 | |
| 42 | 83.8 | |
| 43 | 83.7 | |
| 44 | 83.6 | |
| 45 | 83.6 | |
| 46 | 83.5 | |
| 47 | 83.2 | |
| 48 | 83.1 | |
| 49 | 83.1 | |
| 50 | 83.1 |