Claude Opus 5.5 Leads OpenRouter’s Scanned Document OCR Test

How OpenRouter tested OCR accuracy

OpenRouter created eight test pages from text written by its team, allowing every character to be checked. The pages included a receipt, a letter with two deliberate misspellings, a form with a ticked box, a two-column newsletter and a handwritten note. Three more pages were designed to be difficult: a faded thermal receipt, fast cramped handwriting and a low-resolution fax containing 12 serial numbers with look-alike characters such as 0/O, 1/I/L, 5/S and 8/B.

The company rendered each page from known text and degraded it with rotation, blur, noise and JPEG compression. It then tested 15 vision models, Mistral OCR on OpenRouter and Tesseract. Each model was evaluated using character error rate and 264 key-value checks across three runs. Those values included totals, dates, card digits, serial numbers and deliberate misspellings.

Claude Opus 5.5 topped the results

According to OpenRouter, anthropic/claude-opus-5.5 led the test with 262 of 264 key values correct. Its character error rate across all pages was 0.02%, and the listed cost was $16.42 per 1,000 pages. Mistral OCR followed with 258 of 264 correct, a 0.14% character error rate and a cost of $2.02 per 1,000 pages.

Google’s google/gemini-3.8-flash and several other models reached 254 of 264. The lowest-cost result described by OpenRouter came from inclusionai/ling-3.0-flash-vl with reasoning turned off, which recorded 254 correct values at $0.04 per 1,000 pages. Tesseract 5.5, run locally, recorded 46 of 88 key values and a 26.06% character error rate across all pages.

On the five ordinary pages, every vision model had a character error rate below 0.3%, and all 15 transcribed the neat handwriting perfectly. Most of the difference came from the faded receipt. Claude Opus 5.5 and Mistral OCR got all 33 checks on that page, while GPT-6 Luna got 16.

Why checking OCR output still matters

OpenRouter found that vision models can produce plausible details that do not appear in a scan. In one GPT-6 Luna run, a receipt’s card ending changed from 9026 to 9027, the name changed from PRIYA to SHEVA, and the model added a different Wi-Fi password. The source receipt also had no NET line, but the transcript included one. GPT-6 Luna changed the server’s name and the year in all three runs.

Turning reasoning off made four of five tested models faster and cheaper, with no key values lost. OpenRouter also found that comparing two inexpensive models from different labs could identify errors: flagging disagreements caught 34 of 35 wrong words for $0.62 per 1,000 pages.

The results are a starting point rather than a universal ranking. OpenRouter says its eight pages may not represent a user’s documents, so organizations should test candidate models on their own scans, particularly when processing receipts, serial numbers, IDs or other values that cannot be safely guessed.