Allen Institute's olmOCR Beats GPT-4o at PDF Parsing for a Fraction of the Cost

Researchers at the Allen Institute for AI (AI2) have developed olmOCR, a 7-billion-parameter vision-language model designed to convert complex PDF documents into clean, readable plain text and structured formats like Markdown and LaTeX. The model uses a technique called Document-Anchoring, which combines a PDF page's visual image with extracted backend text coordinates to guide the AI, significantly reducing errors such as hallucinated or fabricated text. To evaluate performance, the team built olmOCR-Bench, a rigorous test suite comprising over 7,010 cases drawn from 1,400 real-world pages including mathematical papers, scanned archival documents, and dense data tables. olmOCR outperformed commercial models including GPT-4o and Gemini Flash 2 across multiple benchmark categories. It also offers a steep cost advantage, processing one million pages for approximately $176 compared to roughly $6,240 for GPT-4o — about 35 times cheaper.