METATR - Handwritten: leaderboard
METATR (2026) tests text recognition systems, from OCR engines to large multimodal models, on diverse real-world documents from archives and libraries in 29 languages with many scripts and layouts. This board is the character error rate in percent on the handwritten documents, which can exceed 100 when a system inserts or hallucinates text; lower is better.
Metric: Character error rate (%) on the handwritten documents, METATR v1.0 (documents from public collections in 29 languages and many scripts and layouts), page-level transcription with the generic instruction (or the recommended OCR prompt of an OCR-tuned model), default hyperparameters, standardized text normalization; can exceed 100 when the output inserts text; lower is better. Source: arxiv.org. Saturation forecast: Around March 2027. 16 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro | 28.5 | #145 |
| 2 | Claude Opus 4.5 | 31.2 | #160 |
| 3 | Mistral OCR | 43.6 | |
| 4 | Azure Document Intelligence Layout | 49.1 | |
| 5 | Qwen 3 VL 8B | 54.9 | #917 |
| 6 | Churro-3B | 58.1 | |
| 7 | GPT-5.1 | 59.5 | #240 |
| 8 | Mistral Medium 3 | 60 | #1096 |
| 9 | Mistral Large 3 | 64.3 | #832 |
| 10 | PeroOCR | 73.9 | |
| 11 | RolmOCR | 89.6 | |
| 12 | olmOCR-2 | 95.3 | |
| 13 | DeepSeek-OCR-2 | 104.4 | |
| 14 | DeepSeek-OCR | 234.5 | |
| 15 | LightOnOCR-2-1B | 255.7 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=metatr-handwritten · How It Works · Data refreshed daily, snapshot 2026-11-21.