Every enterprise AI team eventually hits the same bottleneck: getting useful data out of PDFs. The documents are messy, the tables are weird, the layouts are inconsistent, and the tools that actually work tend to charge like they know you’re desperate. Cohere thinks it has a better deal.
The company released Parse 5 on Thursday, a 2.3-billion-parameter vision language model designed to turn PDFs, slides, and images into structured Markdown. The pitch isn’t that it’s the most accurate parser on the market. It’s that it’s accurate enough while being cheap enough to actually run at scale.
What Parse 5 actually does
Parse 5 (model ID: parse-v5.0) weighs in at roughly 4.6 GB with an 8K context window. It’s built to handle the kinds of documents that make AI pipelines choke: complex tables, nested lists, forms, images with captions, multi-column layouts, and page boundaries.
Tables get exported as HTML. Visual elements come with bounding boxes. Reading order is preserved, which matters more than it sounds, because getting paragraphs in the wrong sequence can quietly wreck a downstream retrieval system.
The model supports nine major commercial languages, which checks a box for multinational deployments in finance, insurance, legal, and scientific fields. Cohere is also offering secure deployment options, a nod to the reality that enterprises in those sectors aren’t thrilled about sending sensitive documents through third-party APIs without guardrails.
The numbers that matter
API pricing starts at $1.50 per 1,000 pages. For organizations processing documents at serious volume, Cohere’s Model Vault offers tiered discounts ranging from 23% to 61% depending on utilization.
To put that in concrete terms: Cohere estimates that a 13-million-page-per-month workflow could save roughly $144K by using the Model Vault instead of straight API calls.
Throughput clocks in at 4.5 pages per second on representative hardware. That speed puts it ahead of some open-source alternatives.
On Cohere’s ParseBench benchmark, Parse 5 scores an average of 79.2. Its table parsing hits 87.0, and faithfulness, meaning how accurately the output reflects the source document, lands at 86.6.
For context, Mistral OCR 4 scores 74.5 on the same benchmark. LlamaParse Cost Effective comes in at 78.3. Parse 5 edges both out, though Cohere is upfront that larger, more expensive models still outperform it on raw accuracy.
Where this fits in the enterprise AI stack
Parse 5 plugs into Cohere’s existing ecosystem and integrates with Microsoft Foundry and AWS SageMaker. That interoperability matters because most enterprises aren’t building their AI stacks from scratch. They’re bolting new tools onto existing cloud infrastructure.
The model is also positioned for agentic retrieval pipelines, the increasingly popular architecture where AI agents autonomously search through document repositories to complete tasks. These workflows tend to process massive volumes of documents, making cost-per-page a critical variable that can determine whether a project is economically viable or dead on arrival.