A hands-on benchmark of 14 OCR engines tested against 93 documents of varying difficulty — from clean invoices to handwritten notes, financial tables, and old newspapers. Engines tested include Tesseract, Docling, Marker, GLM-OCR, PaddleOCR-VL, DeepSeek-OCR, MinerU 2.5, Mistral OCR, Qwen3-VL, Gemini Flash 3.1 Lite, Claude Sonnet 4.6, LlamaParse, and AWS Textract. Key findings: Gemini Flash was the best all-rounder, Tesseract remains unbeatable for clean high-volume docs, Mistral OCR is the cost-effective structured option, and specialized small models underperform on out-of-distribution documents. The core takeaway is that OCR is a routing problem — classify documents and match them to the right engine based on cost, accuracy, and failure tolerance rather than paying for expensive structured APIs across the board.
