BURST-VideoCount: leaderboard
Video counting benchmark built for the Molmo2 paper (Allen Institute for AI and University of Washington, 2026) from the ground-truth object tracks of the BURST test set: about 2,200 queries ask how many instances of an object appear in a video. The score is exact-count accuracy in percent; higher is better.
Metric: Exact-count accuracy (%; 2.2k queries from BURST test-set tracks). Source: arxiv.org. Saturation forecast: Around 2031. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Molmo2-8B | 60.8 | #1488 |
| 2 | GPT-5 Mini | 46 | #349 |
| 3 | Gemini 3 Pro | 44 | #135 |
| 4 | GPT-5 | 43.1 | #163 |
| 5 | Claude Sonnet 4.5 | 42.4 | #244 |
| 6 | Gemini 2.5 Pro | 41.6 | #262 |
| 7 | Gemini 2.5 Flash | 38.7 | #452 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=burst-videocount · How It Works · Data refreshed daily, snapshot 2026-11-12.