Cohere Parse 5: A 2.3B Model That Converts Documents to Markdown at $1.50 per 1,000 Pages

Cohere has released parse-v5.0, a 2.3-billion-parameter vision language model built to convert enterprise documents into structured Markdown in a single pass, no separate OCR step required. The API costs $1.50 per 1,000 pages and is generally available today through Cohere's own API, Microsoft Foundry, AWS SageMaker, and single-tenant Model Vault deployments. The model targets document-heavy industries like financial services, insurance, healthcare, and the public sector, and is designed to feed RAG pipelines, claims processing, and contract search workflows.

What happened

Detail	Value
Model name	parse-v5.0
Parameters	2.3 billion
Context window	8,192 tokens
Model footprint	~4.6 GB
API price	$1.50 per 1,000 pages
Model Vault (Medium)	$4.00/hr or $2,500/month
Model Vault (XL)	$7.00/hr or $4,300/month
ParseBench score (3 of 5 dimensions)	79.2
Stable languages	9 (Arabic, English, French, German, Italian, Japanese, Korean, Portuguese, Spanish)
Availability	Generally available, no waitlist

Cohere built parse-v5.0 on its North-Micro-Vision-Instruct architecture. You send a PDF, PPT, or JPEG page as a base64-encoded data URI and get back Markdown that preserves reading order, tables rendered as HTML, form key-value pairs, image descriptions, and bounding box coordinates. The model handles all of that in one inference pass.

Two output modes are available. The default returns a plain Markdown string per page. Setting output_format="blocks" returns typed blocks where each table carries its HTML, its bounding box, and a description. That second mode is what enables citation-level traceability in RAG applications, where you need to point a user back to the exact page location a fact came from.

How the pricing actually works

The $1.50 per 1,000 pages figure is simple for most teams. But there is a crossover point that Cohere does not advertise directly. At $0.0015 per page, a Medium Model Vault instance ($2,500/month) breaks even at roughly 1.67 million pages per month. An XL instance ($4,300/month) breaks even at roughly 2.87 million pages per month. Below those volumes, metered API calls are cheaper. Above them, dedicated capacity saves money, and usually the real reason enterprises move to Model Vault is data residency or air-gap requirements, not cost alone.

Cohere suggests teams below roughly 100,000 pages per month should start with the free trial key and metered calls before committing to anything.

What does the ParseBench score actually mean?

Cohere reports a ParseBench score of 79.2, which it says beats Mistral OCR 4 (74.5), Azure Document Intelligence (74.3), and Databricks AI Parse (72.4). ParseBench is a LlamaIndex benchmark covering about 2,078 human-verified enterprise pages scored across five dimensions: tables, charts, content faithfulness, semantic formatting, and visual grounding.

Cohere’s 79.2 averages only three of those five dimensions. It drops charts and visual grounding, which are the two dimensions where most parsing tools perform worst. On the full five-dimension public leaderboard, the same competing products score considerably lower: Mistral OCR 4 at 60.68, Databricks AI Parse at 60.68, and Azure Document Intelligence at 59.64. Cohere’s parse-v5.0 is not listed on that public leaderboard at all, where LlamaParse Agentic currently leads at 84.88.

The 79.2 is a vendor-reported subset score, not a leaderboard position. That does not make it worthless, but you should run it against your own document types before treating it as a purchasing signal.

Who it is built for

Cohere targets financial services, insurance, healthcare and life sciences, the public sector, telecom, energy, and manufacturing. These are industries where scanned forms and dense tables are the norm rather than the exception. The model supports RAG ingestion, intelligent document processing, claims and invoice pipelines, contract search, and feeding document context to AI agents.

For teams already building AI integration workflows around document-heavy data, parse-v5.0 is a credible drop-in candidate for the ingestion layer. The nine stable languages cover most global enterprise deployments, and zero-shot support is available for others at reduced accuracy.

Our take

The pricing is genuinely competitive for mid-market RAG pipelines. $1.50 per 1,000 pages is low enough that document parsing stops being a line item to worry about at most volumes. The single-pass architecture (no separate OCR stage) also reduces moving parts, which matters for reliability in production.

The benchmark presentation is the thing to watch carefully. Dropping charts and visual grounding from a benchmark that specifically includes them is not a neutral choice. Those are real failure modes for any team parsing earnings decks, insurance forms with diagrams, or regulatory filings with embedded charts. The honest test is to run parse-v5.0 on a sample of your actual documents, especially anything with complex tables or embedded visuals, before signing a contract.

The availability across Microsoft Foundry and AWS SageMaker is a practical plus. It means teams can deploy within their existing cloud environment without routing data through a new vendor endpoint, which often matters more than benchmark position when you are talking to a security team.

If your team is evaluating document parsing as part of a broader automation build, we cover exactly this kind of workflow automation work for clients, and the ingestion layer is almost always where complexity hides. Recent coverage of access control for AI agents is also worth reading if you are wiring parsed documents into an agentic system.

Bottom line: test parse-v5.0 on your actual documents, pay attention to chart and visual grounding quality, and do the page-volume math before choosing between metered API calls and a Model Vault commitment.

Source: Marktechpost