BELLS: leaderboard
BELLS (CeSIA) — leaderboard for LLM safeguards and supervisors, measuring how reliably a guardrail or frontier model flags harmful or jailbroken interactions; mean BELLS score (%).
Metric: BELLS Score (%). Source: centresecuriteia.github.io. Saturation forecast: Estimated already saturated. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4 | 91.7 | #438 |
| 2 | Claude 3.5 Sonnet | 89.3 | #350 |
| 3 | Grok 2 | 87.5 | #452 |
| 4 | Gemini 1.5 Pro | 84.5 | #445 |
| 5 | DeepSeek V3 | 83.2 | #321 |
| 6 | Mistral Large | 82.5 | #710 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=bells · How It Works · Data refreshed daily, snapshot 2026-10-20.