Cybench
Unguided capture-the-flag cybersecurity tasks.
As of 25 Oct 2026, Claude Opus 4.6 leads Cybench on BenchLeader with 93.0%, ahead of Claude Opus 4.5 at 82.0%, across 21 model configurations with a published result.
- Published by
- Cybenchdata via Epoch AI Benchmarking Hub
- Category
- Agents & tools
- Index weight
- 0.5
- Models
- 21
- Data as of
- 25 Oct 2026
- Newest result
- 5 Feb 2026
The publisher has posted no new results on this board since 5 Feb 2026. BenchLeader re-reads it every day, so newer models appear here as soon as they are published.
CC BY 4.0 — Epoch AI, ‘AI Benchmarking Hub’, epoch.ai/benchmarks. Mirrored boards credit their original publishers.
What the test looks like
Cybench takes capture-the-flag challenges from real security competitions, from cryptography to web exploitation, and asks an agent to find the flag with no hints.
How it is scored
Percent of challenges solved unguided, within a fixed budget of steps and time.
What to keep in mind
Challenge difficulty is uneven and success is binary, so a couple of solves move the score. The agent scaffold matters, and some challenges have public write-ups.
- 1Claude Opus 4.693.0%
- 2Claude Opus 4.582.0%
- 3Claude Sonnet 4.560.0%
- 4Grok 443.0%
- 5Claude Opus 4.142.0%
- 6Grok 4.139.0%
- 7Claude Opus 438.0%
- 8Claude Sonnet 435.0%
- 9Grok 4 Fast30.0%
- 10o3-mini (medium)22.5%
- 11Claude 3.7 Sonnet20.0%
- 12GPT-4.517.5%
- 13Claude 3.5 Sonnet17.5%
- 14GPT-4o12.5%
- 15o110.0%
| # | ||||
|---|---|---|---|---|
| 1 |  Claude Opus 4.6 | 93.0% | 62.5 | 2026-02-05 |
| 2 | Claude Opus 4.5 | 82.0% | 57.9 | 2025-11-24 |
| 3 | Claude Sonnet 4.5 | 60.0% | 53.7 | 2025-09-29 |
| 4 |  Grok 4 | 43.0% | 58.0 | 2025-07-09 |
| 5 | Claude Opus 4.1 | 42.0% | 51.5 | 2025-08-05 |
| 6 | Grok 4.1 | 39.0% | 57.2 | 2025-11-17 |
| 7 | Claude Opus 4 | 38.0% | 52.6 | 2025-05-22 |
| 8 | Claude Sonnet 4 | 35.0% | 50.5 | 2025-05-22 |
| 9 | Grok 4 Fast | 30.0% | – | 2025-09-19 |
| 10 |  o3-minimedium | 22.5% | 45.2 | 2025-01-31 |
| 11 | Claude 3.7 Sonnet | 20.0% | 50.0 | 2025-02-24 |
| 12 | GPT-4.5 | 17.5% | 49.7 | 2025-02-27 |
| 13 | Claude 3.5 Sonnet | 17.5% | 46.3 | 2024-06-20 |
| 14 | GPT-4o | 12.5% | 43.9 | 2024-11-20 |
| 15 | o1 | 10.0% | 53.1 | 2024-09-12 |
| 16 | o1-minimedium | 10.0% | 44.4 | 2024-09-12 |
| 17 | Claude 3 Opus | 10.0% | 39.2 | 2024-02-29 |
| 18 |  Llama 3.1 405Bopen ↗ | 7.5% | 42.9 | 2024-07-23 |
| 19 |  Mixtral 8x22B Instructopen ↗ | 7.5% | 37.4 | 2024-04-17 |
| 20 |  Gemini 1.5 Pro 001 Feb24 | 7.5% | – | 2024-02-15 |
| 21 | Llama 3-70Bopen | 5.0% | 38.6 | 2024-04-18 |
More agents & tools benchmarks · All LLM benchmarks
Cite as: BenchLeader, “Cybench leaderboard”, data as of 25 Oct 2026, https://www.benchleader.com/benchmarks/cybench.
Cybench: questions
- What does Cybench measure?
- Cybench takes capture-the-flag challenges from real security competitions, from cryptography to web exploitation, and asks an agent to find the flag with no hints. Scores are reported in percent of tasks solved; higher is better.
- Which AI model leads Cybench?
- Claude Opus 4.6 leads Cybench with 93.0% as of 25 Oct 2026, ahead of Claude Opus 4.5 at 82.0%.
- How many models have Cybench results?
- 21 model configurations have a Cybench result on BenchLeader, all taken from Cybench via Epoch AI Benchmarking Hub.
- Who runs Cybench and how often is it updated?
- Cybench is published by Cybench. BenchLeader re-reads the published results once a day and records the date each result was published.
- Does Cybench count toward the BenchLeader Index?
- Yes. Cybench contributes to the agents & tools category of the BenchLeader Index, normalised so that 50 is the average of the evaluated models.