CyberBench v1.1 Overall
Agents
CyberBench v1.1 averages the 60-task proof-of-concept and 56-task patch success rates. Vals reran the benchmark using offline task sandboxes and a hardened clean-room grader; these scores are explicitly incomparable with v1. The unequal track sizes mean the aggregate has no single binomial denominator.
Top models (higher is better)
| Model | Score |
|---|---|
| Muse Wonder | 80.4 |
| GPT-6 Sol | 78.0 |
| Gemini 4 Argon | 77.9 |
| GPT-5.6 Sol | 76.3 |
| GPT-6 Luna | 76.3 |
| Claude Haiku 5.5 | 75.5 |
| MiMo-V2.6-Flash | 75.4 |
| Kimi K3 | 75.2 |
| DeepSeek V4.1 Flash | 73.7 |
| GPT-5.6 Luna | 73.6 |
| Muse Wonder Full | 73.6 |
| DeepSeek V4 Pro 0813 | 72.9 |
| MiMo-V2.6-Pro | 72.9 |
| Muse Spark 1.3 | 72.7 |
| DeepSeek-V4-Flash-0731 | 72.1 |
| GLM-5.3 | 72.1 |
| GPT-5.6 Terra | 72.1 |
| Grok 4.5 | 71.0 |
| Ember-1 | 69.5 |
| Muse Spark 1.2 | 69.5 |
| Grok 4.7 | 69.5 |
| GLM-5.3 Flash | 67.9 |
| Qwen3.8 27B | 67.8 |
| Grok 4.6 | 67.7 |
| Inkling | 66.8 |
| Hy4 Preview | 65.4 |
| MiniMax M3 | 65.0 |
| Inkling-Small | 64.1 |
| Sonnet 5 | 61.9 |
| Gemini 3.5 Flash-Lite | 59.1 |
| Mercury 2.5 | 45.5 |
| Gemini 3.6 Flash | 45.4 |
| Gemini 3.7 Flash | 43.8 |
| Gemini 3.8 Flash | 43.8 |
| GPT-6 Astra | 41.1 |
| GPT-6.1 Sol | 39.3 |
| Nemotron 3.5 Lightning 30B A3B BF16 | 29.5 |