Gemini 2.5 Flash OCR returns inconsistent results with identical inputs

I'm using Gemini 2.5 Flash for OCR on PDF documents. The issue is that scanning the same image twice gives different results – first scan is correct, second scan is incorrect. Any suggestions appreciated!

Example:
Correct value: 4589429906422
– First scan: 4589429906422 ✓
– Second scan: 4589429908422 ✗ (misreads "06" as "08")