TraceAV-Bench - Visual Counting: leaderboard

The Visual Counting sub-task of TraceAV-Bench (2026), part of its visual-centric reasoning dimension: multiple-choice questions over long videos whose answers chain several timestamped evidence hops. The score is accuracy in percent with full audio and video input; higher is better.

Metric: Accuracy (%) on the 226 Visual Counting (VC) questions, four-option multiple choice over long audio-visual videos (10 to 140 minutes), exact match of the selected option set, each model at its official inference setting, with the full audio and video input; higher is better. Source: arxiv.org. Saturation forecast: Around 2033. 15 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 3.1 Pro (Preview)41.15#99
2Gemini 2.5 Pro39.38#276
3Gemini 2.5 Flash39.38#463
4Ming-Flash-Omni-2.039.38#1266
5video-SALMONN 239.38#2687
6VideoLLaMA2.1-7B-AV38.94#4204
7MiniCPM-o-4.538.5#1252
8Qwen 3 Omni 30B A3B38.5#847
9Gemini 3 Flash36.73#175
10Qwen2.5-Omni-7B35.4#1884
11HumanOmni-7B34.96#3308
12OmniVinci-9B34.51
13Gemma 4 E4B34.07#2184
14Gemini 2.0 Flash27.43#723
15Baichuan-Omni-1.526.99#3285

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

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