HarmVideoBench - Beyond-Clip Reasoning: leaderboard
Metric: Beyond-Clip Reasoning accuracy (%) on four-option multiple-choice questions about 1,379 real harmful videos (one question per video per level), temperature 0 greedy decoding; API models take the video and audio natively, open models 12 sampled key frames plus the transcript; chance 25; questions that need context beyond the clip; higher is better. Source: arxiv.org. Saturation forecast: Around 2030. 19 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 65.8 |
| 2 | GPT-5.5 | 63.9 |
| 3 | Claude Opus 4.7 | 63.2 |
| 4 | Claude Sonnet 4.6 | 62.7 |
| 5 | Qwen 3 VL 235B A22B (Thinking) | 62.5 |
| 6 | Qwen 3 VL 32B (Thinking) | 62.1 |
| 7 | Qwen 3 VL 235B A22B Instruct | 62 |
| 8 | Qwen 3 VL 30B A3B (Thinking) | 61.2 |
| 9 | Qwen 3 VL 32B Instruct | 61.1 |
| 10 | Qwen 3 VL 8B (Thinking) | 59.5 |
| 11 | Qwen 3 VL 30B A3B Instruct | 58.5 |
| 12 | LFM2.5-VL-1.6B | 51.4 |
| 13 | GLM-4.5V (Non-reasoning) | 50.9 |
| 14 | Qwen 3 VL 8B Instruct | 48.1 |
Interactive version: theaggregate.ai/benchmark?slug=harmvideobench-beyond-clip-reasoning · How It Works · Data refreshed daily, snapshot 2026-09-29.