LVBench
Multimodal
Long-video understanding benchmark with 1,549 four-option questions on 103 videos, covering entity recognition, event understanding, information retrieval, temporal grounding, reasoning and summarization. Scores are multiple-choice accuracy; video sampling budgets are reported per run.
Top models (higher is better)
| Model | Score |
|---|---|
| Gemini 4 Argon | 91.7 |
| GPT-6 Astra | 87.5 |
| Gemini 3.8 Flash | 87.1 |
| Gemini 3.7 Flash | 85.4 |
| Claude Opus 5.5 | 83.7 |
| Claude Fable 5.1 | 79.7 |
| Qwen3.8-Flash-Next | 76.6 |