Vals AI CyberBench: leaderboard
Vals CyberBench v1.1 evaluates agents finding and patching security bugs using mini-swe-agent in offline repository sandboxes: 60 proof-of-concept tasks and 56 reliably graded patch tasks. Overall accuracy combines the two tracks. Provider refusals count as failures; marked fallback-affected results include the disclosed responding models. The hardened clean-room grader makes v1.1 non-comparable with historical v1.
Metric: Accuracy (%). Source: www.vals.ai. Saturation forecast: Not forecast. 47 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Muse Wonder | 80.36 | #15 |
| 2 | GPT-6 Sol | 77.98 | #14 |
| 3 | Gemini 4 Argon | 77.86 | #8 |
| 4 | GPT-5.6 Sol | 76.31 | #19 |
| 5 | GPT-6 Luna | 76.25 | #56 |
| 6 | Claude Haiku 5.5 | 75.48 | #40 |
| 7 | MiMo-V2.6-Flash | 75.36 | #70 |
| 8 | Kimi K3 | 75.24 | #30 |
| 9 | Claude Opus 5.5 | 74.58 | #3 |
| 10 | DeepSeek V4.1 Flash | 73.69 | #36 |
| 11 | GPT-5.6 Luna | 73.63 | #68 |
| 12 | Muse Wonder Full | 73.63 | #10 |
| 13 | MiMo-V2.6-Pro | 72.86 | #42 |
| 14 | DeepSeek V4 Pro (0813) | 72.86 | #53 |
| 15 | Muse Spark 1.3 (Max) | 72.74 | #24 (Muse Spark 1.3) |
Interactive version: theaggregate.ai/benchmark?slug=vals-ai-cyberbench · How It Works · Data refreshed daily, snapshot 2026-10-20.