GitTaskBench (SWE-Agent): leaderboard
GitTaskBench (Ni et al., 2025) gives a code agent a real-world task, such as restoring an old photo or extracting data from a PDF, plus the GitHub repository it must set up and use to do it: 54 tasks from 18 repositories across 7 domains. This board is the task pass rate under SWE-agent: the percentage of tasks whose output meets the task's success criteria, averaged over two runs; higher is better.
Metric: Task pass rate (%). Source: gittaskbench.github.io. Saturation forecast: Estimated already saturated. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude 3.7 Sonnet (20250219) | 42.59 | #399 |
| 2 | GPT-4.1 | 31.48 | #492 |
| 3 | Claude 3.5 Sonnet (20241022) | 22.23 | #635 |
| 4 | O3 Mini (Medium) | 20.37 | #536 (O3 Mini) |
| 5 | Llama 3.3 70B Instruct | 18.52 | #1353 |
| 6 | DeepSeek V3 (0324) | 12.04 | #704 |
| 7 | Qwen 3 32B (Thinking) | 11.11 | #960 (Qwen 3 32B) |
| 8 | GPT-4o (2024-08-06) | 10.19 | #731 |
| 9 | Qwen 3 32B (Non-reasoning) | 3.7 | #960 (Qwen 3 32B) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=gittaskbench-swe-agent · How It Works · Data refreshed daily, snapshot 2026-11-12.