An OpenRouter test of scanned-document transcription found that the most accurate system was not the most economical one, and that plausible invented text remains a serious risk.
The company created eight pages from known text, degraded them with rotation, blur, noise and compression, then tested 15 vision models, Mistral OCR and Tesseract. Claude Opus 5.5 correctly captured 262 of 264 key-value checks at a measured cost of $16.42 per 1,000 pages. Mistral OCR recorded 258 correct checks for $2.02 per 1,000 pages.
Ordinary pages produced little separation: every vision model stayed below a 0.3% character error rate on five simpler scans. Hard cases drove the differences, especially a faded receipt and a fax filled with look-alike characters. Some models also invented realistic-looking names, passwords or values that were absent from the source.
The benchmark is small and uses synthetic pages, so it does not establish a universal winner. Document mix, token accounting and model settings can change both accuracy and price. OpenRouter recommends testing against representative scans, checking completion status, and using disagreement between models to flag uncertain characters rather than trusting a fluent transcript by default.