FAVOR-Bench: leaderboard
Fine-grained video-motion understanding on 8,184 closed-ended multiple-choice questions spanning six motion tasks. Reports ALL answer accuracy (%) across the complete question set, with source video-input settings preserved; separate from action-sequence accuracy and open-ended caption metrics.
Metric: Close-ended ALL accuracy (%). Source: favor-bench.github.io. Saturation forecast: Estimated already saturated. 21 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 1.5 Pro | 49.87 | #445 |
| 2 | Claude 3.7 Sonnet | 43.73 | #249 |
| 3 | GPT-4o | 42.09 | #343 |
| 4 | Qwen 2.5 VL 7B | 40.76 | #643 |
| 5 | InternVL2.5-78B | 38.54 | #431 |
| 6 | InternVL2.5-2B | 22.9 | #1249 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=favor-bench · How It Works · Data refreshed daily, snapshot 2026-10-20.