OpenSkillRisk: leaderboard
Overall Fsafe, the harmonic mean of risk avoidance and explicit risk awareness over 263 risky third-party skills. Codex, Claude Code and Gemini CLI runs remain separate model-and-harness observations; default CLI inference settings.
Metric: Overall Fsafe (%). Source: arxiv.org. Saturation forecast: Around January 2027. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 80.97 | #90 |
| 2 | Claude Opus 4.6 | 80.54 | #63 |
| 3 | GPT-5.4 | 78.46 | #79 |
| 4 | GLM-5.1 | 76.94 | #97 |
| 5 | GPT-5.3 Codex | 73.46 | #69 |
| 6 | Gemini 3.1 Pro (Preview) | 73 | #58 |
| 7 | Kimi K2.6 | 62.69 | #104 |
| 8 | Claude Haiku 4.5 | 59.78 | #280 |
| 9 | DeepSeek V4 Pro | 58.63 | #102 |
| 10 | DeepSeek V3.2 | 53.53 | #205 |
| 11 | Gemini 3 Flash | 49.69 | #99 |
| 12 | GPT-5.1 Codex Mini | 44.42 | #286 |
| 13 | Gemini 2.5 Pro | 27.86 | #152 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=openskillrisk · How It Works · Data refreshed daily, snapshot 2026-10-20.