SWE-bench Verified: leaderboard
The 'Verified' subset of SWE-bench - 500 human-validated GitHub issues ensuring correct problem statements and test coverage. The most trusted variant for measuring real-world software engineering ability.
Metric: Resolved (%). Source: www.swebench.com. Status: saturation imminent. 47 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude Opus 4.5 (High) | 76.8 |
| 2 | Gemini 3 Flash (High) | 75.8 |
| 3 | Claude Opus 4.6 | 75.6 |
| 4 | GPT-5.2 (High) | 72.8 |
| 5 | GPT-5.2 Codex | 72.8 |
| 6 | Claude Sonnet 4.5 (High) | 71.4 |
| 7 | Claude Sonnet 4.5 | 70.6 |
| 8 | Gemini 3 Pro | 69.6 |
| 9 | GPT-5.2 | 69 |
| 10 | Claude Opus 4 (20250514) | 67.6 |
| 11 | GPT-5 (Medium) | 65 |
| 12 | Claude Sonnet 4 (20250514) | 64.9 |
| 13 | Kimi K2 (Thinking) | 63.4 |
| 14 | MiniMax-M2 | 61 |
| 15 | O3 (2025-04-16) | 58.4 |
Interactive version: theaggregate.ai/benchmark?slug=swe-bench-verified · How It Works · Data refreshed daily, snapshot 2026-09-05.