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

Cohere released Parse 5 on Thursday, a document-parsing model aimed at enterprises that need to convert PDFs, slides, and scanned images into structured Markdown for AI pipelines. The company is positioning the product on price-to-performance rather than raw accuracy, a deliberate strategy for organizations processing documents at scale. Parse 5 is a 2.3-billion-parameter vision language model built on Cohere Labs North-Micro-Vision-Instruct architecture, with an 8,192-token context window and a footprint of roughly 4.6 gigabytes. It accepts a page as a base64-encoded image and returns Markdown in reading order, with tables rendered as HTML and bounding box coordinates for tables and images.

Cohere’s own benchmark results show Parse 5 trailing three larger general-purpose models on accuracy. On the ParseBench scoring across tables, content faithfulness, and semantic formatting, Parse 5 scored 79.2, behind GPT-5.5 at 84.4, Opus 4.8 at 84.3, and Gemini 3.5 Flash at 81.8. It outperformed specialized competitors including LlamaParse’s Cost Effective tier at 78.3, Mistral OCR 4 at 74.5, Databricks AI Parse at 72.4, and Azure Document Intelligence at 69.3. Cohere excluded the Layout and Chart dimensions from its table, attributing the omissions to product scope: the model returns reading-order Markdown rather than per-element bounding boxes for text, and it describes charts instead of extracting their underlying data, with chart-data extraction planned for a future version.

The company priced Parse 5 at $1.50 per 1,000 pages through its API, with Model Vault, Cohere’s secure single-tenant platform for managed inference, available for higher-volume deployments. Nils Reimers, VP of AI Search at Cohere, said that document parsing remains unsolved because the hard part is preserving structure and meaning, not reading text. He noted that enterprise documents mix tables, diagrams, charts, and formatting that change interpretation, and that most tools still drop structure or hallucinate content, with even frontier models breaking on layout-heavy pages. On charts specifically, Reimers said Parse 5 provides a general description with an indicator for how agentic AI can visually inspect the chart, rather than extracting data that might miss critical visual cues like line color or pattern.

Cost is the core of Cohere’s sales pitch. Reimers described a modeled workflow for a large financial services firm processing 750 million documents a year, claiming that choosing Parse 5 over a large general-purpose model like GPT-5.5 would reduce costs by more than 98 percent. Cohere presented that figure as its own estimate for a single modeled workflow, not an audited deployment. The model is generally available now through the Cohere API, Model Vault, Microsoft Foundry, and AWS SageMaker. Language support covers Arabic, English, French, German, Italian, Japanese, Korean, Portuguese, and Spanish with stable accuracy, plus lower-accuracy zero-shot support elsewhere. An output mode called blocks returns typed elements where each table carries its own HTML, bounding box, and description, which Cohere positions as enabling citation-level traceability for agents.

Industry analysts see document parsing as central to enterprise AI adoption. Kevin Petrie, VP of Research at BARC US, said a survey his firm is completing shows document analysis is the number one use case for AI, with 62% adoption among organizations polled, and that documents hold the proprietary context needed to differentiate agentic AI initiatives. He said it remains to be seen how Cohere’s cost-performance stacks up against frontier models, but strategically the company has the right focus. Stephanie Walter, Practice Leader for AI Stack at HyperFRAME Research, said Parse 5 sits in a good spot between legacy OCR and using an expensive frontier model on every page. The competitive field includes general-purpose models like GPT-5.5, Opus 4.8, and Gemini 3.5 Flash at the top of the accuracy comparison, specialized parsers such as Mistral OCR 4 and LlamaParse, open-weight options like Chandra OCR 2 and RedNote’s dots.mocr, and hyperscaler services including AWS Textract, Google Document AI, Azure Document Intelligence, and Databricks AI Parse that compete more on ecosystem convenience than raw parsing quality.

More AI news from TechManNews.