CrossSWE - Overall: leaderboard
Cross-repository software engineering requiring coordinated changes to interdependent real repositories. Joint cross-repository pass@1 (%). CrossSWE171instances across165repositories, native common mini-SWE-agent2.4.6 scaffold; high reasoning effort for every model, max300interactionsteps andUSD10modelcost. Pass@1 requires complete acceptance across every participating repository, using native test checks. Single/multilanguage subsets are distinct reported breakdowns, not unchanged SWE-bench results.
Metric: Joint cross-repository pass@1 (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6.1 Sol (Benchmark Sweep Run: source=CrossSWE; mini-SWE-agent2.4.6; high; max300steps; model-cost-budgetUSD10) | 58.48 | |
| 2 | DeepSeek V4.1 Flash (Benchmark Sweep Run: source=CrossSWE; mini-SWE-agent2.4.6; high; max300steps; model-cost-budgetUSD10) | 52.63 | |
| 3 | Claude Sonnet 5.5 (Benchmark Sweep Run: source=CrossSWE; mini-SWE-agent2.4.6; high; max300steps; model-cost-budgetUSD10) | 50.29 | |
| 4 | Qwen3.8-Flash-Next (Benchmark Sweep Run: source=CrossSWE; mini-SWE-agent2.4.6; high; max300steps; model-cost-budgetUSD10) | 45.03 | |
| 5 | GLM-5.3 Flash (Benchmark Sweep Run: source=CrossSWE; mini-SWE-agent2.4.6; high; max300steps; model-cost-budgetUSD10) | 43.86 |
Interactive version: theaggregate.ai/benchmark?slug=crossswe-overall · How It Works · Data refreshed daily, snapshot 2026-11-06.