VRBench: leaderboard
Long narrative video benchmark for multi-step reasoning, assessing both final outcomes and process-level reasoning chains over temporally grounded questions.
Metric: Overall (self-reported). Source: benchmarklist.com. Status: saturated. 29 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | GPT-4o | 68.68 |
| 2 | Claude 3.7 Sonnet | 68.15 |
| 3 | InternVL2.5-78B | 62.31 |
| 4 | Gemini 2.0 Flash (Thinking) | 60.97 |
| 5 | O1 Preview | 60.14 |
| 6 | DeepSeek R1 | 57.13 |
| 7 | DeepSeek V3 | 56.06 |
| 8 | Qwen 2 VL 7B | 54.08 |
| 9 | Qwen 2.5 72B Instruct | 53.51 |
| 10 | QwQ-32B | 52.52 |
| 11 | Llama 3.3 70B Instruct | 49.84 |
| 12 | Qwen 2.5 7B Instruct | 48.29 |
| 13 | QwQ 32B-Preview | 35.9 |
Interactive version: theaggregate.ai/benchmark?slug=vrbench · How It Works · Data refreshed daily, snapshot 2026-09-05.