CAIS Risk Index: leaderboard
Composite risk index imported by benchmarklist.com from the CAIS AI dashboard on 2026-08-26: the mean of five risk scores covering bioweapons assistance (VCT refusal), overconfidence (HLE miscalibration), deception (MASK, Machiavelli) and harmful propensities (TextQuests harm), for models that have all five. It tracks safety behaviour, not task skill, and lower is better. Source: dashboard.safe.ai.
Metric: Risk Index (imported). Source: dashboard.safe.ai. Source: benchmarklist.com. Saturation forecast: Around March 2028. 55 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5.5 | 27.1 | #3 |
| 2 | Claude Fable 5 | 27.3 | #16 |
| 3 | Claude Fable 5.1 | 29 | #9 |
| 4 | Muse Spark 1.1 | 34.2 | #52 |
| 5 | GLM-5.3 | 38.7 | #39 |
| 6 | Claude Opus 4.7 | 39 | #49 |
| 7 | Claude Sonnet 4.5 | 39.6 | #142 |
| 8 | Claude Opus 4.8 | 39.6 | #37 |
| 9 | GLM-5.2 | 42.4 | #73 |
| 10 | Claude Opus 4.5 | 42.7 | #83 |
| 11 | Claude Opus 5 | 44 | #12 |
| 12 | Grok 4.20 | 44.5 | #175 |
| 13 | Claude Sonnet 4.6 | 44.6 | #90 |
| 14 | Grok 4.3 | 44.7 | #188 |
| 15 | GPT-6 Astra | 45.8 | #6 |
No result here: #7 GPT-6.1 Sol, #10 Muse Wonder Full.
Interactive version: theaggregate.ai/benchmark?slug=cais-risk-index · How It Works · Data refreshed daily, snapshot 2026-10-20.