Vision-capable LLMs vs. OCR for long-document (including charts, images, tables, etc.) QA [D]

In three linesBenchmark on 30 long PDFs (171 questions) comparing native vision-LLMs vs OCR pipelines for document QA. Claude Sonnet 4.5 used. LlamaCloud premium achieves 59.6% accuracy ($0.1885/query), native vision 52% ($0.2552/query, most expensive). Vision underperforms on charts/tables; premium OCR more robust. Vision-LLM has 7% intrinsic failure rate vs 0% for OCR after retries.

Summary generated by Claude — human-verified