RealDocBench: leaderboard
RealDocBench (Extend, 2026) tests document parsing on real regulated documents such as mortgage disclosures, financial statements, bills of lading and clinical forms: each system turns a document into markdown, and a fixed extraction model must then answer field-level questions from that markdown alone. The board is per-field accuracy, the percentage of answer fields whose value matches the gold value; higher is better.
Metric: Per-field accuracy (%): share of the 3,742 answer keys whose extracted value matches the typed gold value; RealDocBench QA track: 1,356 field-level questions on 581 real regulated documents (mortgage, finance, supply chain, medical); the system parses each document to markdown without seeing the questions, and a fixed extraction LLM (Gemini 3 Flash) fills the typed answer template from that markdown only; type-tolerant matching; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.8 | 93 | #56 |
| 2 | Gemini 3.5 Flash | 89.3 | #101 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=realdocbench · How It Works · Data refreshed daily, snapshot 2026-11-13.