WildTableBench - Color-based Counting: leaderboard

A WildTableBench (2026) subtype on counting the cells or records that carry a target color, asked about real-world table images from the web. Accuracy in percent, free-form answers judged against the reference by GPT-5.2; higher is better.

Metric: Accuracy (%) on the Color-based Counting subtype (C5-C), 928 human-verified questions over 402 real-world table images, free-form answers judged by GPT-5.2 against the reference answer, high reasoning effort where configurable; higher is better. Source: arxiv.org. Saturation forecast: Around March 2027. 21 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Gemini 3 Pro62.5#146
2Gemini 3 Flash37.5#180
3Seed 2.0 ProHigh37.5#226
4GPT-5.2High30#190
5Kimi K2.5Thinking27.5#272
6Qwen 3 VL 32B (Thinking)25#577
7Claude Opus 4.6High25#114
8Qwen 3 VL 235B A22B (Thinking)22.5#462
9Qwen 3 VL 32B Instruct20#607
10GLM-4.6V17.5#721
11Claude Sonnet 4.6High17.5#179
12Qwen 3 VL 235B A22B Instruct15#541
13GPT-5 MiniHigh15#373
14Qwen 3 VL 4B (Thinking)12.5#1354
15Qwen 3 VL 8B (Thinking)10#924

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=wildtablebench-color-based-counting · How It Works · Data refreshed daily, snapshot 2026-11-22.