What it reads well — and what it doesn't
Text recognition (OCR) is at its best on clear, printed, upright English: documents, receipts, slides, book pages and screenshots at a readable size. It struggles with handwriting, blurry or low-resolution photos, pages photographed at an angle or on a curve, decorative fonts, and tables or multi-column layouts, where the reading order has to be guessed.
ToolGPT is upfront about this. Each job reports how confident the recognition was, and when confidence is low the assistant tells you the text is unreliable instead of presenting it as a clean transcript. It never describes what a page says until the recognised text has actually come back.