Unstructured released upgraded Transform models for its document‑parsing service. The company tested the models on 224 real‑world business pages and found they produced up to 2.3 × fewer table errors than GPT‑6, Claude and Gemini. The tests also showed the competing models hallucinated about 50 % more content that never appeared in the source documents. Unstructured says the new models are ready for customers who need reliable data extraction from PDFs, scans and complex forms.
Why it matters
Businesses that rely on automatic data extraction can expect more accurate tables and fewer hallucinated entries, reducing manual correction work.
