93%
▲ 75.5 pts since 2024
of Cybench tasks solved unguided by Claude Opus 4.6, early 2026
20242026
What the chart shows
Cybench tests AI agents on 40 professional-level capture-the-flag tasks in areas such as cryptography, web security, reverse engineering, forensics and exploitation; the first human solvers took from two minutes to almost 25 hours. The best unguided solve rate rose from 17.5% (Claude 3.5 Sonnet) in 2024 to 82% (Claude 4.5 Opus) in 2025 and 93% (Claude Opus 4.6) in early 2026.
The AI Index calls this the steepest improvement among its agent benchmarks, and suggests capture-the-flag tasks suit current agents well.
Key numbers
| Measure | Value |
|---|---|
| 2024: Claude 3.5 Sonnet | 17.5% |
| 2025: Claude 4.5 Opus | 82% |
| 2026: Claude Opus 4.6 | 93% |
| Change, 2024 to 2026 | +75.5 percentage points |
How to read it
- Unguided means the agent gets each task whole, without it being broken into subtasks.
- The report text says the rate rose from 15% in 2024; the data shows 17.5%, which we use.
- 2026 covers results to early 2026.
Source: Cybench Leaderboard, 2026, via the AI Index Report 2026 (Stanford HAI), figure 2.6.5: Cybench: unguided % solved. Chart and text by AI Adoption.
Numbers checked against the published data on 20 October 2026.
Download the data (CSV) · Cite this chart
Numbers checked against the published data on 20 October 2026.
Download the data (CSV) · Cite this chart
Discussion