Skip to content

Benchmark

ExploitGym · Solved Tasks

Higher scores are better.

Models measured
5
Scores recorded
10
Sources
1
IQ weight
No weight published

Score distribution

131.8
Mean
130
Median
93.6
Std dev
26
10th percentile
247
90th percentile
14
Min
14293
The band covers the middle 80% of scores. The line marks the median.

Leaderboard

#ModelScore
1GPT-5.6 Sol6-hour normalized budget· Reported by Z.ai· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.293
2Claude Fable 56-hour normalized budget· Reported by Z.ai· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.247
3GPT-5.6 Sol2-hour normalized budget· Reported by Z.ai· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.216
4Claude Fable 52-hour normalized budget· Reported by Z.ai· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.181
5GLM 5.36-hour normalized budget· Reported by Z.ai· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.130
6GLM 5.32-hour normalized budget· Reported by Z.ai· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.105
7Kimi K36-hour normalized budget· Reported by Z.ai· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.70
8Kimi K32-hour normalized budget· Reported by Z.ai· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.36
9Qwen3.8 Max6-hour normalized budget· Reported by Z.ai· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.26
10Qwen3.8 Max2-hour normalized budget· Reported by Z.ai· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.14

Data credits: BenchBoard by SYuan03 (github.com/SYuan03/benchboard). MIT covers its code and data organization; no results licence is inferred. Numeric relays retain the original stating-source links.