HarmVideoBench - Observable Evidence: leaderboard
HarmVideoBench (2026) asks multimodal models four-option questions about 1,379 real harmful videos at three levels of understanding; API models watch the video with audio and open models see 12 key frames with the transcript. This board is accuracy in percent on questions about what is directly visible or audible; chance is 25 and higher is better.
Metric: Observable Evidence 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 about what is directly visible or audible in the clip; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 19 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 96 | #94 |
| 2 | GPT-5.5 | 95.4 | #49 |
| 3 | Qwen 3 VL 235B A22B (Thinking) | 95.1 | #451 |
| 4 | Qwen 3 VL 235B A22B Instruct | 94.8 | #522 |
| 5 | Claude Opus 4.7 | 94.7 | #79 |
| 6 | Qwen 3 VL 32B (Thinking) | 94.7 | #559 |
| 7 | Claude Sonnet 4.6 | 94.4 | #161 |
| 8 | Qwen 3 VL 32B Instruct | 94.3 | #585 |
| 9 | Qwen 3 VL 30B A3B (Thinking) | 94 | #712 |
| 10 | Qwen 3 VL 30B A3B Instruct | 91.5 | #871 |
| 11 | Qwen 3 VL 8B (Thinking) | 91.2 | #887 |
| 12 | MiniCPM-V-4.6 | 83.9 | #3525 |
| 13 | LFM2.5-VL-1.6B | 79.8 | #3872 |
| 14 | GLM-4.5V (Non-reasoning) | 77.7 | #744 (GLM-4.5V) |
| 15 | Qwen 3 VL 8B Instruct | 73.3 | #981 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=harmvideobench-observable-evidence · How It Works · Data refreshed daily, snapshot 2026-11-13.