Editorial illustration for Cohere's Parse 5 Model Converts Documents to Markdown, Scores 79.2 on ParseBench
Cohere's Parse 5 Converts PDFs to Markdown at $1.50/1K Pages
Cohere's Parse 5 Model Converts Documents to Markdown, Scores 79.2 on ParseBench
Cohere put a price tag on document parsing this week: $1.50 per 1,000 pages. The company released Parse (parse-v5.0), a 2.3-billion-parameter vision language model built to convert PDFs, PowerPoint slides and JPEG scans into structured Markdown, no separate OCR step required. It runs on an 8,192-token context window, weighs in around 4.6GB, and is built on Cohere Labs' North-Micro-Vision-Instruct architecture.
Feed it a base64-encoded page and Parse hands back text in reading order, tables as HTML, form fields as key-value pairs, and bounding box coordinates for images. That output format matters for the buyers Cohere is chasing: insurance companies drowning in scanned claims forms, banks with decades of dense financial tables, hospitals and life sciences firms with unstructured intake documents. Telecom, energy and manufacturing round out the target list, industries where the paperwork is heavy and the tables rarely behave.
Parse is already generally available through Cohere's own API, Microsoft Foundry, AWS SageMaker and a single-tenant Model Vault option for companies that need data to stay in-house. No waitlist, no research license gating access. Cohere is framing this squarely as a price-performance play, backed by a self-reported benchmark score worth a closer look.
Parse takes a PDF, PPT or JPEG page as a base64-encoded data URI and returns Markdown containing text in reading order, tables rendered as HTML, lists, form key-value pairs, image descriptions and bounding box coordinates.
Why this matters
A 2.3B model beating Azure Document Intelligence and Mistral OCR on a benchmark is worth a second look, but the number to watch is 79.2, not 100. That's roughly one in five documents still getting mangled on tables, faithfulness or formatting, which matters a lot if you're feeding this into a RAG pipeline or a finance workflow where a misread cell or dropped key-value pair costs real money. The 4.6GB footprint and 8,192-token window make Parse cheap to run at scale, and that's the actual pitch here: not accuracy for its own sake, but accuracy-per-dollar for enterprises processing millions of pages.
Worth remembering ParseBench comes from LlamaIndex, not an independent third party, and Cohere is the one reporting its own score against competitors. That's not disqualifying, but it means teams evaluating Parse should run their own messy PDFs and scanned forms through it before trusting the marketing number. For builders choosing between Parse, Mistral OCR and Databricks AI Parse, the gap between 79.2 and 72.4 is real but narrow enough that your specific document type will decide the winner, not the leaderboard.
Related Reading
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
From Archives Google's Gemini Flash AI Extends Video Scenes Up to 40 Seconds
From Archives OpenAI to Launch Ads on ChatGPT Free, India Campaigns Start at Rs 725 Daily
Common Questions Answered
What is Cohere's Parse 5 model and what document formats can it convert?
Cohere's Parse 5 (parse-v5.0) is a 2.3-billion-parameter vision language model designed to convert PDFs, PowerPoint slides, and JPEG scans into structured Markdown format. The model requires no separate OCR step and can process base64-encoded pages to return text in reading order, tables as HTML, lists, form key-value pairs, image descriptions, and bounding box coordinates.
What is the pricing model for Cohere's Parse 5 and what are its technical specifications?
Cohere priced Parse 5 at $1.50 per 1,000 pages, making it cost-effective for large-scale document processing. The model has a 4.6GB footprint, operates on an 8,192-token context window, and is built on Cohere Labs' North-Micro-Vision-Instruct architecture.
How does Parse 5's 79.2 ParseBench score impact its reliability for enterprise workflows?
Parse 5 achieved a 79.2 score on ParseBench, which means approximately one in five documents still experience issues with tables, faithfulness, or formatting. This limitation is significant for critical applications like RAG pipelines or finance workflows where misread cells or dropped key-value pairs can result in costly errors.
How does Parse 5 compare to other document intelligence solutions like Azure Document Intelligence?
Parse 5's 2.3-billion-parameter model outperforms both Azure Document Intelligence and Mistral OCR on the ParseBench benchmark. However, its 79.2 score indicates room for improvement compared to these competitors, particularly in handling complex table structures and maintaining formatting accuracy.
Further Reading
- Introducing Parse: Enterprise document intelligence at scale - Cohere
- Cohere Releases Parse 5 (parse-v5.0): A 2.3B Vision Language Model That Turns Enterprise Documents Into Markdown - MarkTechPost
- ParseBench — Document Parsing Benchmark for AI Agents - ParseBench
- Parse (Details and Application) - Cohere Documentation - Cohere Docs
- Release Notes - Cohere Docs
