OCR-Bench-V2 (EN)Multimodal

The English version of OCRBench v2 was released in early 2025, adding text localization, handwriting understanding, and logical reasoning; it has four times as many task types as v1. It includes more than 10,000 human-verified question-answer pairs across 31 scenarios and a private 1,500-image test set to discourage overfitting, scored out of 100.

Ranking Sources

#ModelScore

1

Qwen3.8-MaxAlibaba

74.2%

2

Qwen3.7-PlusAlibaba

70.7%

3

GPT-5.6 SolOpenAI

69.0%

4

Qwen3.6-PlusAlibaba

67.0%

5

Claude Fable 5Anthropic

65.3%

6

Gemini 3.1 ProGoogle

64.6%

7

Claude Opus 4.6Anthropic

54.3%

8

Claude Opus 4.8Anthropic

53.9%

8 models total

Back to benchmarks