WildTableBench - Cell-Level: leaderboard

Accuracy in percent on the cell-level category: transcribing ranges, reading a cell address, semantic lookup and structure, and writing Excel formulas, free-form questions about real-world table images collected from the web (WildTableBench, 2026), answers judged against the reference by GPT-5.2. Higher is better.

Metric: Accuracy (%) on the Cell-Level question category (C1, question-weighted over its subtypes), 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 Pro62.8#146
2Seed 2.0 ProHigh59.3#226
3GPT-5.2High53.1#190
4Claude Opus 4.6High51.7#114
5Gemini 3 Flash51#180
6Kimi K2.5Thinking47.6#272
7Qwen 3 VL 235B A22B (Thinking)37.9#462
8Claude Sonnet 4.6High35.9#179
9Qwen 3 VL 32B (Thinking)28.3#577
10Qwen 3 VL 32B Instruct24.8#607
11Qwen 3 VL 235B A22B Instruct22.8#541
12GPT-5 MiniHigh20.7#373
13GLM-4.6V17.9#721
14O3High17.2#242
15Qwen 3 VL 8B Instruct10.3#1043

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

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