Cohere has unveiled its document-parsing AI model Parse 5, VentureBeat reported on Aug. 27.

Parse 5 is a 2.3-billion-parameter vision-language model that converts PDFs, slides and images into structured markdown, the report said. Cohere explained it focused on cost performance rather than accuracy.

On its in-house benchmark, ParseBench, Parse 5 scored an average 79.2 points across three categories: tables, content fidelity and semantic formatting. It trailed GPT-5.5 (84.4 points), Opus 4.8 (84.3 points) and Gemini 3.5 Flash (81.8 points), but scored higher than LlamaParse, Mistral OCR (Mistral OCR 4), Databricks AI Parse and Azure Document Intelligence. The API price is $1.50 per 1,000 pages. A single-tenant platform called Model Vault is also offered for companies handling large volumes.

Parse 5 recognizes each page as an image and returns results with a single run of a vision-language model. It integrates an approach that previously handled OCR and the model separately.

Nils Reimers (닐스 라이머스), vice president of Cohere's AI search unit, said, "The reason document parsing is hard is that it lies in preserving structure and meaning." He said, "Most tools miss the structure, and even frontier models make errors on pages with complex layouts. In simulations of financial services workflows processing 750 million documents a year, using Parse 5 instead of GPT-5.5 could cut costs by more than 98 percent."