benchgap
General

ExploitBench leaderboard

As of 2026-11-21, the highest measured score on ExploitBench is 100.0% by GPT-6.2 Astra. 97 more models have estimated scores, calibrated from the benchmarks they were measured on.

Measured scores: benchlm.ai.

#ModelScoreSource
1Claude Opus 5.5100.0%estimated ± 1.9 pp, low confidence
2Claude Sonnet 5.5100.0%estimated ± 1.9 pp, low confidence
3Gemini 4 Carbon100.0%estimated ± 1.9 pp, low confidence
4GPT-6.2 Astra100.0%measured
5GPT-6.2 Sol100.0%measured
6GPT-6 Astra100.0%measured
7GPT-6 Cyber100.0%measured
8Gemini 4 Argon99.8%estimated ± 1.9 pp, medium confidence
9GPT-6.1 Sol99.7%measured
10Claude Fable 5.299.1%estimated ± 6.5 pp, low confidence
11Gemini 4 Neon99.1%estimated ± 6.5 pp, medium confidence
12Claude Fable 5.199.1%estimated ± 6.5 pp, medium confidence
13Claude Opus 598.7%estimated ± 1.9 pp, medium confidence
14GPT-6.1 Terra97.1%measured
15Claude Mythos 5.295.0%measured
16GPT-6 Sol91.7%estimated ± 1.9 pp, medium confidence
17GPT-5.5 Pro84.5%estimated ± 13.4 pp, low confidence
18Gemini 4 Hydrogen83.3%estimated ± 6.5 pp, medium confidence
19GPT-5.4 Pro78.5%estimated ± 13.4 pp, low confidence
20Claude Mythos 578.0%measured
21Kimi K3.1 Pro74.7%measured
22GPT-5.6 Sol73.5%measured
23Claude Mythos Preview69.0%measured
24Gemma 4 31B62.4%estimated ± 7.3 pp, low confidence
25Qwen3.8 Max59.8%estimated ± 7.2 pp, medium confidence
26DeepSeek V4 Flash 073155.3%estimated ± 7.2 pp, medium confidence
27Muse Wonder55.0%estimated ± 6.5 pp, medium confidence
28GLM-5.354.4%measured
29GPT-5.6 Terra52.9%measured
30Gemini 3.5 Flash52.9%estimated ± 7.2 pp, medium confidence
31Qwen3.7 Max52.9%estimated ± 7.2 pp, medium confidence
32Qwen3.8-27B52.3%estimated ± 12.6 pp, low confidence
33Gemma 4 26B A4B50.2%estimated ± 7.3 pp, low confidence
34Gemini 3.8 Flash49.7%estimated ± 1.9 pp, medium confidence
35o349.1%estimated ± 7.2 pp, medium confidence
36MiMo-V2.6-Pro47.9%measured
37Muse Spark 1.343.9%estimated ± 7.2 pp, medium confidence
38Inkling 2 Small42.4%estimated ± 10.4 pp, low confidence
39GPT-541.1%estimated ± 7.2 pp, medium confidence
40Claude Opus 4.841.0%estimated ± 1.9 pp, medium confidence
41Qwen3.6-35B-A3B39.6%estimated ± 7.2 pp, medium confidence
42Muse Spark 1.137.7%estimated ± 13.4 pp, low confidence
43Claude Sonnet 537.6%estimated ± 1.9 pp, medium confidence
44MiniMax M3.1 Flash36.8%estimated ± 12.6 pp, low confidence
45Nemotron 3 Ultra36.7%estimated ± 7.2 pp, low confidence
46Qwen3.5 Flash36.7%estimated ± 7.2 pp, low confidence
47Qwen 3.6 Max (preview)35.2%estimated ± 7.2 pp, low confidence
48GPT-5.535.0%estimated ± 1.9 pp, medium confidence
49SWE-2 Flash34.8%measured
50Muse Spark Open34.3%estimated ± 12.6 pp, low confidence
51Hy4 preview33.9%estimated ± 12.6 pp, low confidence
52GPT-6 Luna33.7%estimated ± 1.9 pp, medium confidence
53Qwen3.5 397B33.6%estimated ± 7.2 pp, low confidence
54Claude Haiku 5.533.2%estimated ± 6.5 pp, medium confidence
55GPT-5.6 Luna33.2%measured
56Grok 4.632.4%estimated ± 6.5 pp, medium confidence
57GPT-5.232.4%estimated ± 6.5 pp, medium confidence
58GPT-5.432.4%estimated ± 6.5 pp, medium confidence
59Claude Fable 532.4%estimated ± 6.5 pp, medium confidence
60DeepSeek V4.1 Flash 101732.4%estimated ± 6.5 pp, medium confidence
61Claude Opus 4.732.4%estimated ± 6.5 pp, low confidence
62Claude Opus 4.532.4%estimated ± 6.5 pp, low confidence
63Claude Opus 4.632.4%estimated ± 6.5 pp, low confidence
64Gemini 3.7 Flash32.4%estimated ± 6.5 pp, low confidence
65Gemini 3.6 Flash32.4%estimated ± 6.5 pp, low confidence
66Grok 4.532.4%estimated ± 6.5 pp, low confidence
67Kimi K2.632.4%estimated ± 6.5 pp, low confidence
68Grok 4.732.4%estimated ± 6.5 pp, low confidence
69Mistral Large 432.4%estimated ± 6.5 pp, low confidence
70Claude Sonnet 4.532.1%estimated ± 7.2 pp, low confidence
71Qwen3.5-122B-A10B32.1%estimated ± 7.2 pp, low confidence
72Qwen3.5 Plus32.1%estimated ± 7.2 pp, low confidence
73Qwen3.7 Plus32.1%estimated ± 7.2 pp, low confidence
74Kimi K332.0%measured
75DeepSeek V4 Pro31.9%estimated ± 12.6 pp, low confidence
76Gemini 3.1 Pro31.5%estimated ± 1.9 pp, low confidence
77Gemini 3.5 Flash-Lite31.5%estimated ± 1.9 pp, low confidence
78Gemini 3 Flash31.5%estimated ± 1.9 pp, low confidence
79Ling 3.1 Flash31.3%estimated ± 12.6 pp, low confidence
80Grok 4.2031.1%estimated ± 13.4 pp, low confidence
81Claude Sonnet 4.630.5%estimated ± 7.2 pp, low confidence
82GPT-5.130.5%estimated ± 7.2 pp, low confidence
83GLM-5.228.8%estimated ± 7.2 pp, low confidence
84Qwen3.7 Flash28.8%estimated ± 7.2 pp, low confidence
85Muse Spark 1.228.6%estimated ± 13.4 pp, low confidence
86Seed 2.1 Turbo27.4%estimated ± 13.4 pp, low confidence
87Step 5 Preview26.1%estimated ± 12.6 pp, low confidence
88MiMo-V2.6-Flash25.3%measured
89GPT-4o mini23.8%estimated ± 7.2 pp, low confidence
90Qwen3.6 Plus23.8%estimated ± 7.2 pp, low confidence
91GLM-4.721.4%estimated ± 13.4 pp, low confidence
92Inkling21.4%estimated ± 13.4 pp, low confidence
93Seed 2.1 Pro21.4%estimated ± 13.4 pp, low confidence
94GLM-5.120.3%estimated ± 13.4 pp, low confidence
95Grok 4.318.0%estimated ± 13.4 pp, low confidence
96Beam 1.0 501B-A23B17.9%estimated ± 12.6 pp, low confidence
97GLM-5.3-Flash16.5%estimated ± 7.2 pp, low confidence
98MiniMax M316.5%estimated ± 7.2 pp, low confidence
99GPT-4.1 mini14.6%estimated ± 7.2 pp, low confidence
100GPT-5.4 mini14.6%estimated ± 7.2 pp, low confidence
101o3-mini14.6%estimated ± 7.2 pp, low confidence
102Qwen3.6-27B14.6%estimated ± 7.2 pp, low confidence
103Kimi K2.7 Code13.7%estimated ± 13.4 pp, low confidence
104Inkling-Small12.6%estimated ± 7.2 pp, low confidence
105GPT-5.4 nano10.6%estimated ± 7.2 pp, low confidence
106o4-mini10.6%estimated ± 7.2 pp, low confidence
107Qwen3 Max10.6%estimated ± 7.2 pp, low confidence
108GPT-4.110.6%estimated ± 13.4 pp, low confidence
109o19.6%estimated ± 13.4 pp, low confidence
110GPT-4o8.6%estimated ± 13.4 pp, low confidence
111Claude Haiku 4.57.6%estimated ± 13.4 pp, low confidence
112MAI-Thinking-15.6%estimated ± 12.6 pp, low confidence
113Claude 3 Opus3.7%estimated ± 13.4 pp, low confidence
114GPT-OSS 120B0.0%estimated ± 7.2 pp, low confidence
115GPT-4.1 nano0.0%estimated ± 13.4 pp, low confidence

All leaderboards

Agentic · terminal

Agentic · tools

Coding

Math

Knowledge & reasoning

Instruction following

Multilingual

Vision & documents

Long context

General