WildTableBench - Color: leaderboard
Metric: Accuracy (%) on the Color question category (C5, 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 April 2028. 21 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3 Pro | 55.8 |
| 2 | Gemini 3 Flash | 37.5 |
| 3 | Kimi K2.5 (Thinking) | 37.5 |
| 4 | Seed 2.0 Pro (High) | 32.5 |
| 5 | GPT-5.2 (High) | 29.2 |
| 6 | Qwen 3 VL 32B (Thinking) | 28.3 |
| 7 | Qwen 3 VL 235B A22B (Thinking) | 26.7 |
| 8 | GLM-4.6V | 25.8 |
| 9 | Qwen 3 VL 235B A22B Instruct | 21.7 |
| 10 | Claude Opus 4.6 (High) | 20.8 |
| 11 | Qwen 3 VL 32B Instruct | 18.3 |
| 12 | GPT-5 Mini (High) | 18.3 |
| 13 | Claude Sonnet 4.6 (High) | 17.5 |
| 14 | O3 (High) | 15 |
| 15 | Qwen 3 VL 8B (Thinking) | 11.7 |
Interactive version: theaggregate.ai/benchmark?slug=wildtablebench-color · How It Works · Data refreshed daily, snapshot 2026-10-07.