AIR-BENCH Live: leaderboard
AIR-BENCH Live (2026) is a self-updating successor to AIR-BENCH 2024 that adds risk categories from new AI regulation and regenerates persona-driven attack prompts, one base and one authority-endorsement mutation per level-4 category; a GPT-5.4-mini judge scores refusal 1, partial compliance 0.5 and full compliance 0. Score is the mean safety score times 100 over 2,680 prompts in English, Spanish, Japanese and Portuguese; higher is safer.
Metric: Mean safety score (x100; %; 2,680 prompts in English, Spanish, Japanese and Portuguese) on regenerated persona-driven attack prompts over 335 level-4 risk categories (one base and one authority-endorsement mutation per category), scored by a GPT-5.4-mini judge as 1 for refusal, 0.5 for partial or evasive, 0 for full compliance. Source: arxiv.org. Saturation forecast: Not forecast. 14 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Haiku 4.5 | 89 | #568 |
| 2 | Grok 4.3 | 72 | #333 |
| 3 | Llama 3.3 70B | 64 | #1498 |
| 4 | Gemini 2.5 Pro | 60 | #266 |
| 5 | Qwen 3 235B A22B | 60 | #623 |
| 6 | Kimi K2 | 58 | #439 |
| 7 | GPT-4o | 57 | #741 |
| 8 | Gemini 2.5 Flash | 53 | #453 |
| 9 | Llama 3 8B | 53 | #5637 |
| 10 | DeepSeek V3.2 | 52 | #406 |
| 11 | DeepSeek R1 | 48 | #473 |
| 12 | Mistral Large | 17 | #2074 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=air-bench-live · How It Works · Data refreshed daily, snapshot 2026-11-13.