Optical character recognition has come a long way from the noisy outputs of early commercial products. In 2026, even commodity engines achieve character error rates (CER) below 2% on clean printed text. But the interesting edge cases—handwriting, degraded scans, multi-column layouts, and embedded tables—separate the leaders from the pack.
Our benchmark corpus spanned 14,000 pages across five document types: typed business letters, scanned legal contracts, government forms, academic papers with LaTeX equations, and mixed handwriting/print field reports.
Methodology: We normalized page resolution to 300 DPI, applied no pre-processing beyond deskew, and scored CER against manually verified ground truth. Table extraction was scored on a cell-level F1 metric.
Results summary: Cloud Vision API led on clean printed text (CER 0.4%) but fell to 11.2% CER on degraded scans. Tesseract 5 with LSTM showed the most balanced profile across document types. The newest commercial entrant outperformed all others on table extraction (F1 0.91) but added significant per-page cost.
Practical recommendation: For high-volume, low-complexity documents, open-source Tesseract with pre-processing delivers acceptable quality at zero marginal cost. For contracts and compliance documents where extraction errors carry legal risk, the premium cloud engines justify their pricing.