Cohere Releases Document Analysis Model "Parse 5"

Cohere has publicly released "Parse 5," a document analysis model that converts PDFs, slides, and images into Markdown format. The API usage price is set at $1.50 per 1,000 pages, and the company is positioning cost-performance balance rather than accuracy as its key value proposition for enterprise customers.

When enterprises attempt to feed large volumes of document data into AI systems, the main barrier has traditionally been a binary choice between "accuracy or cost." Using high-precision models increases processing costs, while cheaper tools often lose table and graph layout structures or generate non-existent content. Parse 5 is positioned as a model that addresses this challenge with the direction of "sufficient accuracy at an affordable price."

Parse 5 has 2.3 billion parameters and adopts the "North-Micro-Vision-Instruct" architecture developed by Cohere Labs. The context window is 8,192 tokens, and the model file size is approximately 4.6GB. Input is received as Base64-encoded images, supporting PDF, PowerPoint, and JPEG formats. Output is primarily Markdown strings per page, with tables in HTML format and images annotated with captions and bounding box coordinates. The model maintains stable accuracy in English, Japanese, French, German, Spanish, Portuguese, Italian, Korean, and Arabic, while also functioning in other languages with reduced accuracy.

The distinctive feature lies in its processing method. Most traditional tools follow a two-stage process: first extracting text from documents using OCR (optical character recognition), then analyzing meaning with a separate model. Parse 5 adopts a "single-pass" approach, processing everything with a single vision-language model, which simplifies the workflow and reduces costs. Cohere explains that the bounding boxes and HTML-formatted table data in the output enhance traceability when AI agents cite or reference specific information sources.

According to Cohere's published benchmark comparison, Parse 5 scores 79.2 (combined score of tables, content fidelity, and semantic formatting). This falls below GPT-5.5's 84.4, Opus 4.8's 84.3, and Gemini 3.5 Flash's 81.8, while exceeding LlamaParse (score not specified in original). The company does not claim the top score but rather emphasizes that it "achieves near-top accuracy at the lowest cost." Nils Reimers, VP of AI Search at Cohere, stated that "the challenge in document analysis is not reading text, but preserving structure and meaning," noting that even general-purpose large models can fail on pages with complex layouts.

Parse 5 is currently available through Cohere API, the company's secure single-tenant platform "Model Vault," Microsoft Foundry, and AWS SageMaker. For large-scale processing, deployment via Model Vault is recommended.

Rather than competing solely on accuracy, this approach of transparently pricing and clearly showing "what accuracy level you can get at what cost" makes it easier for enterprises to structure their decision-making when integrating large document volumes into AI pipelines. While large language models continue to improve in accuracy, processing costs often become a practical constraint in real enterprise deployments. It will be worth watching how widely such cost-competitive specialized models become adopted.