WildTableBench - Row Operation: leaderboard
Metric: Accuracy (%) on the Row Operation subtype (C4-R), 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 January 2027. 21 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3 Pro | 77.5 |
| 2 | Claude Opus 4.6 (High) | 67.5 |
| 3 | Gemini 3 Flash | 57.5 |
| 4 | Kimi K2.5 (Thinking) | 57.5 |
| 5 | Claude Sonnet 4.6 (High) | 50 |
| 6 | Seed 2.0 Pro (High) | 47.5 |
| 7 | GPT-5.2 (High) | 45 |
| 8 | GPT-5 Mini (High) | 37.5 |
| 9 | GLM-4.6V | 35 |
| 10 | Qwen 3 VL 32B (Thinking) | 35 |
| 11 | Qwen 3 VL 235B A22B (Thinking) | 30 |
| 12 | Qwen 3 VL 32B Instruct | 27.5 |
| 13 | Qwen 3 VL 235B A22B Instruct | 25 |
| 14 | Qwen 3 VL 8B (Thinking) | 20 |
| 15 | Qwen 3 VL 8B Instruct | 15 |
Interactive version: theaggregate.ai/benchmark?slug=wildtablebench-row-operation · How It Works · Data refreshed daily, snapshot 2026-10-07.