LVBench: leaderboard
LVBench asks multiple choice questions about 103 long public videos that average about an hour, each question having one correct option and three distractors. Its 1,549 items cover temporal grounding, summarization, reasoning, entity recognition, event understanding and key information retrieval. The number is accuracy, higher is better, and these figures are self-reported. Source: arxiv.org/abs/2406.08035.
Metric: Score (imported). Source: arxiv.org. Source: benchmarklist.com. Source: llm-stats.com. Saturation forecast: Around December 2027. 49 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 4 Argon | 91.7 | #8 |
| 2 | Gemini 3.8 Flash | 87.8 | #21 |
| 3 | GPT-6 Astra | 87.5 | #6 |
| 4 | Gemini 3.7 Flash | 85.4 | #23 |
| 5 | Gemini 3.6 Flash | 84.2 | #51 |
| 6 | Claude Opus 5.5 | 83.7 | #3 |
| 7 | GPT-5.6 Sol | 82.1 | #19 |
| 8 | Qwen 3.8 Max | 81.8 | #35 |
| 9 | Claude Fable 5.1 | 79.7 | #9 |
| 10 | GPT-5.6 Terra | 78.9 | #41 |
| 11 | Seed 2.1 Pro | 78 | #67 |
| 12 | GPT-5.4 (xHigh) | 77.4 | #79 (GPT-5.4) |
| 13 | Qwen3.8 Omni Flash | 76.9 | #28 |
| 14 | Seed 2.1 Turbo | 76.8 | #86 |
| 15 | Qwen3.8-Flash-Next | 76.6 | #72 |
No result here: #7 GPT-6.1 Sol, #10 Muse Wonder Full.
Interactive version: theaggregate.ai/benchmark?slug=lvbench · How It Works · Data refreshed daily, snapshot 2026-10-20.