Skip to content

Benchmark

CyberGym Level 1 success rate (agent-focused systems)

Agentic benchmarks measure how well models carry out tasks that take several steps. Higher scores are better.

Models measured
12
Scores recorded
37
Sources
1
IQ weight
0.66

Score distribution

0.888
Mean
0.913
Median
0.097
Std dev
0.731
10th percentile
0.971
90th percentile
0.602
Min
0.6020.994
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 0.66 for chat models, measured over 12 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.

Leaderboard

#ModelScore
1DeepSeek V4.1 Flashagent: TuLongFeng, trials: 1, features: dynamic, reported by: Qihoo 360, date: 2026-09-23· CyberGym leaderboard data docs0.994
2DeepSeek V4.1 Flashagent: Crimson, trials: 1, features: dynamic, reported by: Ansen, date: 2026-09-18· CyberGym leaderboard data docs0.99
3DeepSeek V4.1 Flashagent: RO0T Agent, trials: 1, features: dynamic, reported by: RO0T Lab, date: 2026-09-18· CyberGym leaderboard data docs0.977
4GLM 5.3agent: Sangfor AI, trials: 1, features: Orchestration, Multi-stage, dynamic, reported by: Sangfor AI, date: 2026-09-02· CyberGym leaderboard data docs0.972
5DeepSeek V4 Flashagent: Ostorlab Neutron, trials: 1, features: dynamic, reported by: Ostorlab, date: 2026-10-06· CyberGym leaderboard data docs0.971
6DeepSeek V4 Flashagent: VARAS-OneMind (智安·宇盾), trials: 1, features: dynamic, test-time mem., reported by: VARAS, date: 2026-09-16· CyberGym leaderboard data docs0.97
7GLM 5.3agent: Alipay AI4SDL, trials: 1, features: dynamic, reported by: Alipay, date: 2026-09-06· CyberGym leaderboard data docs0.968
8GLM 5.3agent: TieYing (铁鹰), trials: 1, features: dynamic, reported by: China Mobile, date: 2026-10-10· CyberGym leaderboard data docs0.964
9DeepSeek V4 Proagent: Panshi (磐石), trials: 1, features: dynamic, reported by: Huawei Cloud, date: 2026-09-04· CyberGym leaderboard data docs0.954
10GLM 5.3agent: NSFOCUS AI, trials: 1, features: dynamic, reported by: NSFOCUS, date: 2026-08-13· CyberGym leaderboard data docs0.95
11GLM 5.3agent: DoGNAVY, trials: 1, features: Multi-agent, Memory, dynamic, reported by: deepsec@DARKNAVY, date: 2026-08-17· CyberGym leaderboard data docs0.95
12Claude Opus 5agent: AWS Continuum, trials: 1, reported by: AWS, date: 2026-10-15· CyberGym leaderboard data docs0.942
13DeepSeek V4 Flashagent: RO0T Agent, trials: 1, features: dynamic, reported by: RO0T Lab, date: 2026-08-26· CyberGym leaderboard data docs0.942
14GLM 5.3agent: BladeSec Agent, trials: 1, features: dynamic, reported by: BladeSec, date: 2026-10-09· CyberGym leaderboard data docs0.936
15DeepSeek V4 Flashagent: Sangfor AI, trials: 1, features: Orchestration, Multi-stage, dynamic, reported by: Sangfor AI, date: 2026-08-08· CyberGym leaderboard data docs0.932
16DeepSeek V4 Flashagent: PwnBot, trials: 1, features: dynamic, test-time mem., reported by: FUNNY_04, date: 2026-09-27· CyberGym leaderboard data docs0.931
17DeepSeek V4.1 Flashagent: AiPy Mini, trials: 1, features: dynamic, reported by: Knownsec, date: 2026-09-23· CyberGym leaderboard data docs0.928
18DeepSeek V4 Flashagent: NASH, trials: 1, features: dynamic, reported by: NASH Team, date: 2026-09-01· CyberGym leaderboard data docs0.914
19DeepSeek V4 Flashagent: FangcunCyber, trials: 1, features: test-time mem., reported by: Fangcun AI, date: 2026-08-16· CyberGym leaderboard data docs0.913
20DeepSeek V4 Flashagent: Whitzard (白泽), trials: 1, features: dynamic, reported by: Fudan Whitzard, date: 2026-08-07· CyberGym leaderboard data docs0.912
21Claude Opus 4.6agent: mneme, trials: 1, features: dynamic, reported by: Soongsil AISC, date: 2026-07-13· CyberGym leaderboard data docs0.91
22GLM 5.2agent: DoGNAVY, trials: 1, features: Multi-agent, Memory, dynamic, reported by: deepsec@DARKNAVY, date: 2026-08-03· CyberGym leaderboard data docs0.908
23GLM 5.3agent: adversarial-patch-agent-v1, trials: 1, features: dynamic, reported by: Red Hat, date: 2026-10-12· CyberGym leaderboard data docs0.907
24DeepSeek V4 Proagent: HiAUDIT Agent, trials: 1, features: dynamic, test-time mem., reported by: TECHFUND Inc., date: 2026-09-18· CyberGym leaderboard data docs0.906
25Claude Opus 4.6agent: Crystalline, trials: 1, features: test-time mem., Test-time memory, reported by: Independent researcher, date: 2026-06-08· CyberGym leaderboard data docs0.896
26DeepSeek V4 Flashagent: TriLane, trials: 1, features: dynamic, reported by: CMHI (China Mobile (Hangzhou) Information Technology Co., Ltd.), date: 2026-09-20· CyberGym leaderboard data docs0.889
27DeepSeek V4 Flashagent: Gusion, trials: 1, reported by: Gusion, date: 2026-08-16· CyberGym leaderboard data docs0.871
28GLM 5.2agent: Sangfor AI, trials: 1, features: Orchestration, Multi-stage, dynamic, reported by: Sangfor AI, date: 2026-07-21· CyberGym leaderboard data docs0.863
29GLM 5.2agent: RedbudAI, trials: 1, features: dynamic, test-time mem., reported by: Redbud, date: 2026-08-10· CyberGym leaderboard data docs0.863
30GLM 5.2agent: Xuanwu Atuin AI, trials: 1, features: dynamic, reported by: Tencent Xuanwu Lab, date: 2026-07-22· CyberGym leaderboard data docs0.848
31GLM 5.1agent: Xuanwu Atuin AI, trials: 1, features: dynamic, reported by: Tencent Xuanwu Lab, date: 2026-07-02· CyberGym leaderboard data docs0.84
32GPT 5.4agent: VARAS-OneMind (智安·宇盾), trials: 1, features: test-time mem., reported by: VARAS, date: 2026-08-14· CyberGym leaderboard data docs0.738
33MiniMax M3agent: MopMonk Agent, trials: 1, reported by: MopMonk AI, date: 2026-06-28· CyberGym leaderboard data docs0.731
34GLM 5.1agent: JiuXuan (九玄), trials: 1, features: Memory, dynamic, reported by: China Mobile Jiutian AI, date: 2026-08-03· CyberGym leaderboard data docs0.729
35GLM 5.1agent: Whitzard (白泽), trials: 1, features: dynamic, reported by: Fudan Whitzard, date: 2026-07-21· CyberGym leaderboard data docs0.689
36DeepSeek-V4-Pro Previewagent: XD Agent, trials: 1, reported by: XDxAI, date: 2026-08-11· CyberGym leaderboard data docs0.641
37GPT 5agent: SageAgent, trials: 1, features: dynamic, reported by: OpenSage Team, date: 2026-02-09· CyberGym leaderboard data docs0.602

Data credits: CyberGym leaderboard (Wang, Shi, He, Cai, Zhang, Song; UC Berkeley), ICLR 2026. Results evaluated and submitted by the listed teams. No licence is stated for the result data; the repository's Apache-2.0 licence covers its code only.