Cohere Parse 5 loses the benchmark on points. It wins on cost per page.

Cohere's Parse 5 AI model has released a new pricing strategy focused on cost efficiency, rather than solely on raw performance. The 2.3-billion-parameter vision language model is designed to convert PDFs, slides, and images into structured Markdown at scale, addressing the common issue of tools missing document structure and exceeding cost thresholds for large-scale deployment.

While Parse 5 falls behind larger general-purpose models like GPT-5.5, Opus 4.8, and Gemini 3.5 Flash in benchmark tests, Cohere emphasizes its competitive pricing at $1.50 per 1,000 pages. This pricing puts Parse 5 close to the top performers, providing an attractive cost-to-capability ratio for enterprises.

Parse 5's single-pass architecture simplifies the document parsing process by treating each page as an image and running it through a single vision-language model pass. The resulting structured Markdown includes tables as HTML, image descriptions, and bounding box coordinates, preserving the document's structure and meaning.

Cohere's VP of AI Search, Nils Reimers, acknowledges that preserving structure and semantic formatting remains a significant challenge for document parsing. He explains that Parse 5's general description of charts, rather than extracting raw data, helps avoid hallucinations in agentic AI applications. The company has demonstrated significant cost savings for large financial institutions by switching to Parse 5 over GPT-5.5, estimating a more than 98% cost reduction.

The model supports 8 languages (Arabic, English, French, German, Italian, Japanese, Korean, Portuguese, and Spanish) and offers two output modes: markdown per page and typed elements (tables with HTML, bounding boxes, and descriptions). Parse 5 is now available through Cohere's API, Model Vault, Microsoft Foundry, and AWS SageMaker.