Vals CyberBench: leaderboard
Historical CyberBench v1 results mirrored by BenchmarkList. Agents find and patch vulnerabilities in repository sandboxes. Vals superseded this release with v1.1 offline task sandboxes and a hardened clean-room grader; its 60-task PoC and 56-task Patch scores are not directly comparable to v1. The current primary board is Vals AI CyberBench.
Metric: Score (%). Source: benchmarklist.com. Saturation forecast: Not forecast. 20 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol | 88.14 | #22 |
| 2 | GPT-5.6 Luna | 83.9 | #79 |
| 3 | GPT-5.5 | 80.51 | #33 |
| 4 | Kimi K3 | 79.03 | #35 |
| 5 | GLM-5.2 | 77.36 | #80 |
| 6 | MiniMax-M3 | 76.27 | #151 |
| 7 | Kimi K2.6 | 74.35 | #113 |
| 8 | GPT-5.4 (2026-03-05) | 72.88 | #63 |
| 9 | Gemini 3.5 Flash Lite | 72.32 | #174 |
| 10 | Gemini 3.5 Flash | 69.33 | #66 |
| 11 | DeepSeek V4 Pro | 68.64 | #112 |
| 12 | DeepSeek V4 Flash | 66.72 | #155 |
| 13 | Inkling | 64.14 | #144 |
| 14 | Claude Opus 4.7 | 52.14 | #51 |
| 15 | Claude Opus 4.8 | 50.85 | #39 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=vals-cyberbench · How It Works · Data refreshed daily, snapshot 2026-11-05.