VideoZeroBench: leaderboard

VideoZeroBench (2026) asks 500 manually annotated, deliberately hard open-ended questions with exact short answers about 138 long videos in 13 domains, each with annotated temporal intervals and bounding boxes as supporting evidence. Level-3 accuracy, the standard end-to-end setting with only the video and the question, in percent; higher is better.

Metric: Level-3 accuracy (%), standard end-to-end question answering with no evidence hints, on VideoZeroBench's 500 manually annotated open-ended questions (exact short answers) over 138 long videos in 13 domains; Gemini models get the raw video through the API, the others 1 fps frames up to the printed cap, temperature 0; higher is better. Source: arxiv.org. Saturation forecast: Around 2028. 17 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 3 Pro17#145
2Gemini 2.5 Pro16.4#276
3Gemini 2.5 Flash11.4#463
4Qwen 3.5 397B A17B10.6#264
5Seed 2.0 Pro10.6#222
6InternVL3.5-8B9.8#1604
7Qwen 3 VL 235B A22B9.6#485
8VideoChat-R1.5-7B8.6#1254
9Qwen 3 VL 8B8.2#917
10GPT-5.28#185
11Qwen 3 VL 4B7.8#1151
12Video-R17.8#1379
13Qwen 2.5 VL 7B6#1862
14InternVL3.5-4B5.8#1930
15VideoRFT-7B5.8#1863

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=videozerobench · How It Works · Data refreshed daily, snapshot 2026-11-21.