XL-DocBench (Image): leaderboard
The same 1,519 expert-verified questions over documents of up to 2,303 pages, answered from rendered page images instead of extracted text. Every model evaluated both ways scores lower here than on the OCR board, so the gap between the two is the cost of reading a document as pixels.
Metric: Accuracy (%). Source: officeintelligence.github.io. Saturation forecast: Around 2029. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 26.9 | #79 |
| 2 | Kimi K2.5 | 25.4 | #144 |
| 3 | Qwen 3.5 35B A3B | 24.2 | #256 |
| 4 | GPT-5.2 | 22.9 | #110 |
| 5 | Qwen 3.5 9B | 19.8 | #379 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=xl-docbench-image · How It Works · Data refreshed daily, snapshot 2026-10-20.