Cybench

Unguided capture-the-flag cybersecurity tasks.

Published by

Cybenchdata via Epoch AI Benchmarking Hub

Category

Agents & tools

Index weight

0.5

Models

21

Data as of

9 Sept 2026

CC BY 4.0 — Epoch AI, ‘AI Benchmarking Hub’, epoch.ai/benchmarks. Mirrored boards credit their original publishers.

1. 1Claude Opus 4.693.0%
2. 2Claude Opus 4.582.0%
3. 3Claude Sonnet 4.560.0%
4. 4Grok 443.0%
5. 5Claude Opus 4.142.0%
6. 6Grok 4.139.0%
7. 7Claude Opus 438.0%
8. 8Claude Sonnet 435.0%
9. 9Grok 4 Fast30.0%
10. 10o3-mini22.5%
11. 11Claude 3.7 Sonnet20.0%
12. 12GPT-4.517.5%
13. 13Claude 3.5 Sonnet17.5%
14. 14GPT-4o12.5%
15. 15o110.0%

SearchAll providersAnthropicGoogleMetaMistral AIOpenAIxAIOpen weights21 of 21

#	Model	Score ↓	Run ↕	Index ↕	Published ↕
1	Claude Opus 4.6Anthropic	93.0%	–	63.1	–
2	Claude Opus 4.5Anthropic	82.0%	–	58.0	–
3	Claude Sonnet 4.5Anthropic	60.0%	–	52.3	–
4	Grok 4xAI	43.0%	0709	57.2	–
5	Claude Opus 4.1Anthropic	42.0%	–	52.6	–
6	Grok 4.1xAI	39.0%	–	48.8	–
7	Claude Opus 4Anthropic	38.0%	–	50.7	–
8	Claude Sonnet 4Anthropic	35.0%	–	49.5	–
9	Grok 4 FastxAI	30.0%	–	46.5	–
10	o3-miniOpenAI	22.5%	–	45.6	–
11	Claude 3.7 SonnetAnthropic	20.0%	–	49.8	–
12	GPT-4.5OpenAI	17.5%	–	49.6	–
13	Claude 3.5 SonnetAnthropic	17.5%	–	45.1	–
14	GPT-4oOpenAI	12.5%	–	42.3	–
15	o1OpenAI	10.0%	–	53.2	–
16	o1-miniOpenAI	10.0%	–	44.4	–
17	Claude 3 OpusAnthropic	10.0%	–	38.9	–
18	Llama 3.1 405BMetaopen	7.5%	–	42.6	–
19	Mixtral 8x22BMistral AIopen	7.5%	–	39.0	–
20	Gemini 1.5 Pro 001 Feb24Google	7.5%	–	–	–
21	Llama 3-70BMetaopen	5.0%	–	38.6	–