HarmVideoBench - Observable Evidence: leaderboard

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

#ModelScore
1Gemini 3.1 Pro (Preview)96
2GPT-5.595.4
3Qwen 3 VL 235B A22B (Thinking)95.1
4Qwen 3 VL 235B A22B Instruct94.8
5Claude Opus 4.794.7
6Qwen 3 VL 32B (Thinking)94.7
7Claude Sonnet 4.694.4
8Qwen 3 VL 32B Instruct94.3
9Qwen 3 VL 30B A3B (Thinking)94
10Qwen 3 VL 30B A3B Instruct91.5
11Qwen 3 VL 8B (Thinking)91.2
12LFM2.5-VL-1.6B79.8
13GLM-4.5V (Non-reasoning)77.7
14Qwen 3 VL 8B Instruct73.3

Interactive version: theaggregate.ai/benchmark?slug=harmvideobench-observable-evidence · How It Works · Data refreshed daily, snapshot 2026-09-29.