VEBench - Footage Selection: leaderboard

The footage selection task of VEBench (2026): choose which of four candidate clips best continues a reference video as A-roll and B-roll editing would, with the candidates concatenated and marked on screen and subtitles provided. Accuracy in percent; higher is better.

Metric: Footage Selection accuracy (%) on Operation Simulation: four-option choice of the candidate clip that best fits a reference video (A/B-roll pairing), candidates concatenated into one sequence with on-screen text markers, subtitles provided; higher is better. Source: arxiv.org. Saturation forecast: Around April 2027. 9 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 2.5 Pro44.44#276
2GPT-4o41.89#751
3Qwen 3 VL 32B Instruct39.72#606
4Qwen 3 VL 8B Instruct36.14#1029
5Qwen 3 VL 4B Instruct31.8#1332
6InternVL3-14B31.16#1214
7InternVL3-8B30.78#1782
8Qwen 2.5 VL 7B Instruct30.01#1902
9Qwen 3 VL 30B A3B Instruct29.37#895

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

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