WildTableBench: leaderboard

WildTableBench (2026) asks multimodal models 928 human-verified free-form questions about 402 table images collected from the web, in 17 subtypes and five categories. The score is overall accuracy in percent, judged against the reference answer by GPT-5.2; higher is better.

Metric: Overall accuracy (%) on WildTableBench, 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 Pro67.9#145
2Kimi K2.5Thinking49.9#266
3Gemini 3 Flash49.4#175
4Seed 2.0 ProHigh47.8#222
5GPT-5.2High46.6#185
6Claude Opus 4.6High45.7#113
7Claude Sonnet 4.6High35#173
8Qwen 3 VL 235B A22B (Thinking)34.7#459
9Qwen 3 VL 32B (Thinking)28.2#573
10GPT-5 MiniHigh25.3#364
11Qwen 3 VL 235B A22B Instruct25.2#532
12Qwen 3 VL 32B Instruct24.5#606
13GLM-4.6V24.4#710
14O3High14.8#237
15Qwen 3 VL 8B (Thinking)14.7#910

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

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