SWE-Touch: leaderboard
SWE-Touch — tests whether a coding agent can still finish a repair after the user directly edits the code in the shared workspace mid-task. Counter-Edit resolve rate (%) on SWE-bench Verified: the same tasks as the agent's autonomous (Vanilla) run, but with a conflicting user commit landing during the run, which costs the 9 evaluated models a mean 7.7 points.
Metric: Counter-Edit Resolve Rate (%). Source: trae1oung.github.io. Saturation forecast: Around February 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.8 | 83.3 | #39 |
| 2 | GPT-5.5 | 79.2 | #33 |
| 3 | Qwen 3.7 Max | 70.3 | #85 |
| 4 | GLM-5.1 | 68.3 | #109 |
| 5 | MiniMax-M2.5 | 66.2 | #316 |
| 6 | Kimi K2.6 | 64.3 | #113 |
| 7 | DeepSeek V4 Pro | 63.8 | #112 |
| 8 | MiniMax-M2.7 | 62.7 | #270 |
| 9 | Qwen 3 Coder 480B A35B | 40.7 | #286 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=swe-touch · How It Works · Data refreshed daily, snapshot 2026-11-05.