Cohere released Parse 5 on Thursday: a 2.3-billion-parameter vision language model that converts PDFs, slides, and images into structured Markdown at $1.50 per 1,000 pages. On ParseBench, it scores 79.2, behind GPT-5.5 (84.4), Opus 4.8 (84.3), and Gemini 3.5 Flash (81.8). Cohere is not claiming the top score. It is claiming that for a financial services workflow processing 750 million documents a year, choosing Parse 5 over GPT-5.5 cuts costs by more than 98 percent. That figure is Cohere's own model, not an audited deployment.
The architecture is worth reading in detail. Parse 5 runs a single vision-language pass on each page image, collapsing the separate OCR-plus-model pipeline most tools use. It runs on a 4.6-gigabyte footprint with an 8,192-token context window. Its blocks output mode returns typed elements, each table carrying its own HTML, bounding box, and description, which is the format Cohere positions as the foundation for citation-level traceability in agentic pipelines. Two ParseBench dimensions, Layout and Chart, are excluded from Cohere's published score. Chart data extraction is planned for a future version.
The benchmark gap is the argument for reading the full piece, not against it. Parse 5 sits above LlamaParse Cost Effective (78.3), Mistral OCR 4 (74.5), Databricks AI Parse (72.4), and Azure Document Intelligence (69.3), while costing a fraction of frontier models. BARC US research puts document analysis at 62 percent adoption among polled organizations, the top AI use case by a wide margin. The real question, as HyperFRAME Research's Stephanie Walter frames it, is not whether the parser read the PDF but whether the agent can correctly use what came out of it.
[READ ORIGINAL →]