Cohere’s Parse 5 Turns Enterprise Documents Into Markdown for $1.50 per 1,000 Pages
The compact vision model is built for high-volume retrieval and document-processing pipelines, but its headline benchmark result leaves out chart and visual-grounding tests.
Listen to this story
The audio brief
Story brief
3 key pointsCohere is making Parse 5 generally available as a low-cost document-ingestion option for enterprise RAG pipelines, with API pricing of $1.50 per 1,000 pages. Its 2.3B-parameter model handles PDFs, PowerPoints and JPEGs in one pass, preserving reading order, tables, forms, images and coordinates in Markdown. The tradeoff is incomplete handoff data: no confidence scores or hierarchy labels, Markdown instead of JSON,...
- 01
Parse 5 supports Arabic, English, French, German, Italian, Japanese, Korean, Portuguese and Spanish.
- 02
Cohere reports 79.2 across three ParseBench dimensions; the score excludes charts and visual grounding.
- 03
The model is available through Cohere’s API, Microsoft Foundry, AWS SageMaker and single-tenant Model Vault.
Cohere has released Parse 5, a 2.3-billion-parameter vision-language model that converts enterprise document pages into structured Markdown for $1.50 per 1,000 pages. The product targets a costly step in retrieval-augmented generation, or RAG: turning source documents into material that AI systems can search and use.
Parse 5, also called parse-v5.0, accepts PDF, PowerPoint and JPEG pages as base64-encoded data URIs. It returns reading-order text, lists, form key-value pairs, image descriptions and bounding-box coordinates, alongside tables rendered as HTML within Markdown.
One pass, layout included
The model performs document text recovery without a separate OCR stage before it. Its output is designed to retain enough structure for document search, retrieval workflows and downstream processing: tables keep their HTML representation, while bounding boxes locate content on a page.
Cohere says nine languages are stable: Arabic, English, French, German, Italian, Japanese, Korean, Portuguese and Spanish. The model is generally available through the Cohere Parse API, Microsoft Foundry, AWS SageMaker and the single-tenant Model Vault offering, without a waitlist.
A selective benchmark comparison
Cohere reports a 79.2 ParseBench score averaged over tables, content faithfulness and semantic formatting. On that same three-dimension comparison, Mistral OCR 4 scored 74.5, Azure Document Intelligence 74.3 and Databricks AI Parse 72.4.
ParseBench evaluates five dimensions across about 2,078 human-verified enterprise pages: tables, charts, content faithfulness, semantic formatting and visual grounding. Parse 5’s reported average excludes charts and visual grounding, so it should not be read as an all-dimensions leaderboard position.
The handoff still needs work
- Parse 5 does not provide confidence scores for extracted content.
- It does not identify header or font hierarchy, limiting information available for section-aware processing.
- It produces Markdown rather than structured JSON, and its 8,192-token context window limits processing per invocation.
Those limits define the boundary of Cohere’s pitch. Parse 5 can supply structured page output, but applications that need a fixed schema or a measure of extraction certainty require capabilities the model does not return directly.
Price is the commercial argument
The API price is $1.50 per 1,000 pages. Cohere’s Model Vault option is listed at $2,500 per month for a Medium instance and $4,300 per month for XL, offering a separate single-tenant deployment route from metered API use.
The central question for prospective users is therefore document-specific rather than settled by the 79.2 figure: whether Parse 5’s selective benchmark result, layout-rich Markdown output and stated limitations fit the documents and downstream controls in their own ingestion pipeline.
Sources
- marktechpost.comCohere Releases Parse 5 (parse-v5.0): A 2.3B Vision Language Model That Turns Enterprise Documents Into Markdown