Cybench leaderboard: which LLMs score highest
Cybench (cybersecurity capture-the-flag tasks) currently has scores for 16 models. The top score is 43.0% by grok-4-0709; the median model scores 17.5% and the lowest scores 5.0%.
| # | Model | Organisation | Cybench score | Cheapest input $/M | Source | Captured |
|---|---|---|---|---|---|---|
| 1 | grok-4-0709 | xAI | 43.0% | — | Epoch AI Benchmarking Hub | 2026-10-05 |
| 2 | claude-opus-4-1-20250805 | Anthropic | 42.0% | — | Epoch AI Benchmarking Hub | 2026-10-05 |
| 3 | grok-4-1 | xAI | 39.0% | — | Epoch AI Benchmarking Hub | 2026-10-05 |
| 4 | Claude Opus 4 | Anthropic | 38.0% | $15.00 | Epoch AI Benchmarking Hub | 2026-10-05 |
| 5 | claude-sonnet-4-20250514 | Anthropic | 35.0% | — | Epoch AI Benchmarking Hub | 2026-10-05 |
| 6 | grok-4-fast | xAI | 30.0% | — | Epoch AI Benchmarking Hub | 2026-10-05 |
| 7 | claude-3-7-sonnet-20250219 | Anthropic | 20.0% | — | Epoch AI Benchmarking Hub | 2026-10-05 |
| 8 | Claude 3.5 Sonnet (Jun 2024) | Anthropic | 17.5% | $3.00 | Epoch AI Benchmarking Hub | 2026-10-05 |
| 9 | GPT-4.5 Preview (Feb 2025) | OpenAI | 17.5% | $75.00 | Epoch AI Benchmarking Hub | 2026-10-05 |
| 10 | GPT-4o (Nov 2024) | OpenAI | 12.5% | $2.50 | Epoch AI Benchmarking Hub | 2026-10-05 |
| 11 | claude-3-opus-20240229 | Anthropic | 10.0% | $15.00 | Epoch AI Benchmarking Hub | 2026-10-05 |
| 12 | o1-preview | OpenAI | 10.0% | $15.00 | Epoch AI Benchmarking Hub | 2026-10-05 |
| 13 | gemini-1.5-pro-001-feb24 | Google DeepMind | 7.5% | — | Epoch AI Benchmarking Hub | 2026-10-05 |
| 14 | Llama-3.1-405B-Instruct | Meta AI | 7.5% | $3.50 | Epoch AI Benchmarking Hub | 2026-10-05 |
| 15 | Mixtral-8x22B-Instruct-v0.1 | Mistral AI | 7.5% | $0.600 | Epoch AI Benchmarking Hub | 2026-10-05 |
| 16 | Meta-Llama-3-70B-Instruct | Meta AI | 5.0% | $0.120 | Epoch AI Benchmarking Hub | 2026-10-05 |
Compare all benchmarks side by side on the live leaderboard →
Among the ten highest scorers, the cheapest listed API price belongs to GPT-4o (Nov 2024) at $2.50 per million input tokens (score 12.5%).
What Cybench measures
Cybench measures cybersecurity capture-the-flag tasks.
How to read these scores
Scores are percentages (higher is better): the share of questions or tasks the model completed correctly. Where a model is listed at several reasoning efforts, only its highest-effort row is shown. See the methodology for how rows are chosen.
Where the data comes from
Epoch AI Benchmarking Hub (CC BY 4.0). Scores were last captured 2026-10-05; the Captured column gives each row's date.