ClashBench - Successful Preemption Rate (Claude Code): leaderboard
Coding agents that share infrastructure with an incumbent workload, driven by the Claude Code harness: share of runs in which the agent's interference actually seizes the contested resource. Passive overwrite, lease/lock, name collision, elastic capacity and quota exhaustion conflicts; lower is better. Source: arxiv.org/abs/2609.19892.
Metric: SPR (%). Source: arxiv.org. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 5 | 25.2 | #65 |
| 2 | MiniMax-M3 | 27.2 | #151 |
| 3 | GLM-4.7 | 33 | #202 |
| 4 | GLM-5.2 | 38.3 | #80 |
| 5 | Qwen 3.5 9B | 41 | #391 |
| 6 | Qwen 3.7 Flash | 46.8 | #142 |
| 7 | DeepSeek V4 Flash (0731) | 47.1 | #84 |
| 8 | GLM-5.3 | 47.3 | #44 |
| 9 | Qwen 3.6 35B A3B | 48.5 | #238 |
| 10 | Qwen 3.6 27B | 49 | #182 |
| 11 | Qwen 3.8 27B | 50.8 | #123 |
| 12 | Qwen 3.7 Plus | 55.6 | #98 |
| 13 | Qwen 3.8 Max | 61.2 | #34 |
Interactive version: theaggregate.ai/benchmark?slug=clashbench-successful-preemption-rate-claude-code · How It Works · Data refreshed daily, snapshot 2026-11-05.