safety benchmark
AIR-Bench
AIR-Bench 2024 is a safety benchmark grounded in risk categories derived from government regulations and company policies. It evaluates policy-grounded refusal across a broad regulatory and policy-derived harm taxonomy, using category-specific LLM-judge prompts that reward safe engagement rather than only penalizing unsafe responses.
Updated Aug 7, 2026
Published models1
Registry coverage1
MetricScore
EvidenceB
AIR-Bench leaderboard
Sorted by the source-provided rank. Higher score is better according to the registry.
| 1 | MI | 88.0% | 100.0% | 1 | C |
AIR-Bench highlights
The top published results on this benchmark's own scale.
What is AIR-Bench?
Definition and scoring fields from the benchmark registry.
AIR-Bench 2024 is a safety benchmark grounded in risk categories derived from government regulations and company policies. It evaluates policy-grounded refusal across a broad regulatory and policy-derived harm taxonomy, using category-specific LLM-judge prompts that reward safe engagement rather than only penalizing unsafe responses.
Scores are shown in ratio. The current registry marks this benchmark as not independently verified with evidence level B.
- Family
- AIR-Bench
- Modality
- text
- Primary category
- safety
- Score direction
- higher
- LLMBoard eligible
- No
- Evaluation key
- air-bench|llm-stats-current
Source-native results are preserved. Eligibility for the overall LLMBoard score is a separate policy decision.
FAQ
Common questions about AIR-Bench.
Which model scores highest on AIR-Bench?
MAI-Thinking-1 is currently ranked first with 88.0%.
What does AIR-Bench measure?
AIR-Bench 2024 is a safety benchmark grounded in risk categories derived from government regulations and company policies. It evaluates policy-grounded refusal across a broad regulatory and policy-derived harm taxonomy, using category-specific LLM-judge prompts that reward safe engagement rather than only penalizing unsafe responses.
Is a higher score better?
Yes. Higher values rank better for this benchmark.
How many models are compared?
1 unique published model results are currently shown.
Does this benchmark affect the overall score?
This benchmark is preserved as source-native evidence but is not eligible for the current overall score.