Cohere's Parse 5 Beats Azure and Mistral at 95% Lower Cost

Cohere shipped Parse 5, a compact vision language model built for one job: converting messy enterprise PDFs, slides, and scans into clean Markdown that downstream agents and RAG pipelines can actually work with. The pitch is familiar in document AI, but the numbers behind it are worth a closer look.

A small model with a specific job

Parse 5 is a 2.3B-parameter multimodal model that weighs roughly 4.6GB, with an 8K context window and a proprietary architecture Cohere calls north-micro-vision-instruct. It extracts text in reading order, tables, lists, forms, images and captions, page boundaries, and the locations of visual elements, then returns everything as Markdown with HTML-formatted tables, bounding box coordinates, and image descriptions attached.

Inputs include PDF, PPT, and Base64-encoded JPEG. The model is stable across nine languages (Arabic, English, French, German, Italian, Japanese, Korean, Portuguese, and Spanish) and supports zero-shot parsing of others at somewhat reduced accuracy.

Benchmark numbers on ParseBench

Cohere is anchoring its claims on ParseBench, LlamaIndex's recently released evaluation. It measures how well document parsing tools convert PDFs into structured output that AI agents can reliably act on, testing whether parsed output preserves the structure and meaning needed for autonomous decisions rather than just resembling a reference text. The benchmark is stratified into five capability dimensions (tables, charts, content faithfulness, semantic formatting, and visual grounding) drawn from around 2,000 human-verified pages across insurance, finance, government, and other domains.

On the three subsets Cohere highlights (tables, content faithfulness, and semantic formatting), Parse 5 scores 79.2 overall. That puts it ahead of the specialist competition:

• Cohere Parse 5: 79.2
• Mistral OCR 4: 74.5
• Azure Document Intelligence: 74.3
• Databricks AI Parse: 72.4

One caveat worth flagging: LlamaParse Agentic was the only method LlamaIndex tested that stayed competitive across all five dimensions, scoring 84.9% overall on the full benchmark. Cohere's headline number covers a three-dimension slice, so it is not quite an apples-to-apples top-of-leaderboard claim.

Pricing that makes ingestion pipelines pencil out

Parse 5 lists at $1.50 per 1,000 pages. Cohere frames that as up to 95% cheaper than routing pages through frontier LLMs or hyperscaler document services, and 63% below Mistral's OCR pricing. For a team processing millions of pages a month, that gap compounds fast, which is where the model's real positioning lives: high-volume ingestion where per-page economics dominate accuracy tradeoffs.

What it is actually good for

The docs are explicit about the target workloads. Parse is built for pipelines where throughput and cost dominate, specifically:

• Preparing documents for search and retrieval
• RAG ingestion at scale
• Intelligent document processing (forms, invoices, filings)
• Feeding document context to AI agents

The Markdown-with-bounding-boxes output shape suits chunking pipelines that want to preserve layout hints and cite back to page regions. Tables come out as HTML rather than flattened into whitespace-separated text, which matters for anything doing numeric extraction.

Known rough edges

Cohere's documentation is refreshingly direct about what Parse will not do:

• No confidence scores on extracted content, so you cannot filter low-certainty spans downstream
• No identification of headers, footers, or font hierarchy, which limits section-aware chunking
• No structured JSON output, only Markdown, so schema extraction still needs an LLM pass

The 8K context window also caps how much of a document Parse sees at once, which works page-by-page but means orchestration logic sits outside the model.

How to try it

Parse 5 is live through the Cohere API, Model Vault for single-tenant deployment, AWS SageMaker, and Microsoft Foundry. A free Hugging Face Space lets you kick the tires without an API key. Existing North and Compass customers get access as part of their platform upgrade.

A minimal call looks like this:

import cohere
co = cohere.ClientV2()
result = co.parse(
    model="parse-v5.0",
    file=open("filing.pdf", "rb"),
)
print(result.output)  # Markdown with HTML tables + bboxes

Where this fits in the market

Document parsing has quietly become one of the most contested corners of enterprise AI. Frontier LLMs can parse a PDF but cost too much to run over millions of pages. Hyperscaler services like Azure Document Intelligence are cheaper but less accurate on messy layouts. LlamaParse Agentic sits at the accuracy frontier with its own cost profile. Parse 5 is Cohere's argument that a small, focused vision model in the 2 to 3 billion parameter range can hit a sweet spot: better than the hyperscalers, dramatically cheaper than frontier models, and fast enough to run on modest hardware. Whether that trade holds up on your document mix is worth an afternoon on the free Space before committing.