GameXpert-Bench - GameFix - Strict Repair Score: leaderboard
GameXpert-Bench GameFix (arXiv 2608.21833v1, Table 3): coding agents repair defects injected into 100 game tasks, graded by executable fail-to-pass and pass-to-pass tests over three runs, where a task score is the percentage of injected bugs fixed. Strict is the share of tasks scoring at least t, averaged over runs and thresholds t = 90-100; 0-100, higher is better. Rows keep harness and reasoning effort.
Metric: Share of tasks with at least t% of injected bugs fixed, averaged over three runs and t = 90-100 (0-100). Source: arxiv.org. Saturation forecast: Around August 2028. 17 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5 (Max) | 39 | #16 (Claude Opus 5) |
| 2 | Claude Fable 5 (Max) | 33.7 | #20 (Claude Fable 5) |
| 3 | GPT-5.6 Sol (xHigh) | 26.7 | #22 (GPT-5.6 Sol) |
| 4 | Claude Opus 4.8 (Max) | 18 | #39 (Claude Opus 4.8) |
| 5 | Claude Opus 4.7 (Max) | 17.2 | #51 (Claude Opus 4.7) |
| 6 | GPT-5.5 (xHigh) | 16.1 | #33 (GPT-5.5) |
| 7 | DeepSeek V4 Flash (Max) | 15.2 | #155 (DeepSeek V4 Flash) |
| 8 | Kimi K3 (Max) | 14 | #35 (Kimi K3) |
| 9 | Gemini 3.5 Flash (High) | 12.2 | #66 (Gemini 3.5 Flash) |
| 10 | MiniMax-M3 (Thinking) | 10.3 | #151 (MiniMax-M3) |
| 11 | Hy3 (High) | 6 | #100 (Hy3) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=gamexpert-bench-gamefix-strict-repair-score · How It Works · Data refreshed daily, snapshot 2026-11-05.