Cohere has shipped parse-v5.0, a 2.3-billion-parameter vision language model that turns PDFs, PowerPoint decks, and JPEGs into structured Markdown with HTML tables, bounding boxes, and reading-order text. No separate OCR stage sits in front of it — the model handles everything in a single pass. The footprint is modest at roughly 4.6GB, and the context window clocks in at 8,192 tokens. Enterprise buyers will find it across Microsoft Foundry, AWS SageMaker, and Cohere’s own API, with single-tenant Model Vault deployment for teams that need air-gapped infrastructure.

The pricing story is unusually transparent for the document parsing space. Cohere wants $1.50 per 1,000 pages on the metered API. Dedicated Model Vault instances run $2,500 a month for Medium or $4,300 for XL. Here’s the crossover nobody advertises: at $0.0015 per page, a Medium instance only breaks even around 1.67 million pages per month. XL needs roughly 2.87 million. Below those volumes, you’re better off paying per page. Above them, dedicated capacity wins on price alone — though data residency requirements are usually what actually push enterprises to Vault.

Now the benchmark. Cohere reports a ParseBench score of 79.2, putting it ahead of Mistral OCR 4 at 74.5 and Azure Document Intelligence at 74.3. But that figure averages only three of ParseBench’s five dimensions — tables, content faithfulness, and semantic formatting. The company drops charts and visual grounding, which happen to be the two categories where most parsers fall apart. On the full five-dimension leaderboard, Mistral and Databricks AI Parse both land at 60.68, and Azure at 59.64. Cohere Parse isn’t even listed there. The 79.2 is a vendor-reported subset, not a leaderboard position.

That’s not disqualifying — it’s just a reason to test on your own documents before believing the slide deck. Cohere is positioning Parse on price-performance rather than peak accuracy, and for high-volume RAG ingestion across financial services, insurance, and healthcare, that trade-off may genuinely make sense. The “blocks” output mode, which returns typed blocks with HTML tables and bounding boxes, is the feature that makes citation-level traceability possible — the kind of thing legal and compliance teams will care about more than any benchmark number. Nine languages are listed as stable, with zero-shot support elsewhere at lower accuracy. The real question isn’t whether 79.2 is honest. It’s whether charts and visual grounding matter for your pipeline.