Cohere launches Parse 5 model, prioritizing cost-effectiveness over raw accuracy

Cohere has released Parse 5, a new vision-language model designed to convert PDFs, slides, and images into structured Markdown format for enterprise use. The company is positioning the model based on its price-to-performance ratio rather than pure accuracy, aiming to provide a scalable solution for businesses processing large volumes of documents.

In its own benchmark comparisons, Cohere acknowledges that Parse 5 scores lower than three larger, general-purpose frontier models: GPT-5.5, Opus 4.8, and Gemini 3.5 Flash. However, the company claims that Parse 5 offers a near-top score at a significantly lower cost per page, with an API price of $1.50 per 1,000 pages.

Nils Reimers, VP of AI Search at Cohere, explained that the primary challenge in document parsing is preserving structure and meaning, not just reading text. Enterprise documents often combine tables, diagrams, and formatting that alter data interpretation. He noted that many tools still drop structure or produce inaccurate content, and even leading models can struggle with layout-heavy pages.

Parse 5 utilizes a single-pass architecture where a document page is processed as an image by a unified vision-language model. This approach generates structured Markdown output, including rendered HTML tables, image descriptions, and bounding box coordinates. The model supports multiple languages with stable accuracy, such as Arabic, English, French, and German.

Cohere estimates that implementing Parse 5 in a modeled workflow for a large financial services firm, processing 750 million documents annually, could reduce costs by over 98% compared to using general-purpose models. This highlights the company's strategy to offer cost-effective solutions for businesses handling substantial data volumes.