Benchmarks · safety
CyberGym-E2E-AA Refusal Rate
Share of CyberGym-E2E-AA tasks the model refused on safety grounds. Lower means fewer refusals on defensive security work.
17 models · 17 results · data as of 2026-10-29 · artificialanalysis.ai ↗
Leaderboard
DeepSeek V4.1 Flash 1017 max
All results
17 rows| # | Model | Effort | Score |
|---|---|---|---|
| 1 | DeepSeek V4.1 Flash 1017 | max | 0.0% |
| 2 | DeepSeek V4.1 Flash | max | 0.0% |
| 3 | GLM 5.3 Flash | max | 0.0% |
| 4 | GLM 5.3 | max | 0.0% |
| 5 | GPT-6 Luna | max | 0.0% |
| 6 | Grok 4.7 | xhigh | 0.0% |
| 7 | Kimi K3 | max | 0.0% |
| 8 | MiMo V2.6 Pro | 0.0% | |
| 9 | MiniMax M3 | 0.0% | |
| 10 | Muse Spark 1.3 | xhigh | 0.0% |
| 11 | Gemini 3.8 Flash | high | 86.3% |
| 12 | Opus 5.5 | max | 98.5% |
| 13 | Fable 5.1 | max | 99.2% |
| 14 | GPT-6 Astra | max | 100.0% |
| 15 | GPT-6 Sol | max | 100.0% |
| 16 | Qwen 3.8 2.4T-A95B | 100.0% | |
| 17 | Qwen 3.8 27B | xhigh | 100.0% |