WildTableBench - Conditional Numerical: leaderboard

A WildTableBench (2026) subtype on an aggregate over the rows that satisfy a filter condition, 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 Conditional Numerical subtype (C2-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 February 2027. 21 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Gemini 3 Pro74.1#145
2Gemini 3 Flash54.3#175
3Kimi K2.5Thinking52.6#266
4Seed 2.0 ProHigh52.6#222
5Claude Opus 4.6High47.4#113
6GPT-5.2High42.2#185
7Qwen 3 VL 235B A22B (Thinking)37.9#459
8Claude Sonnet 4.6High37.1#173
9Qwen 3 VL 32B (Thinking)31#573
10Qwen 3 VL 235B A22B Instruct30.2#532
11GLM-4.6V28.4#710
12GPT-5 MiniHigh27.6#364
13Qwen 3 VL 32B Instruct23.3#606
14Qwen 3 VL 8B (Thinking)20.7#910
15Qwen 3 VL 8B Instruct11.2#1029

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

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