Cohere shipped Parse, a vision language model that converts contracts, invoices, and scanned reports into structured Markdown, and priced API access at $1.50 per 1,000 pages. It handles tables and pictures sitting inside a page rather than plain text alone, and covers nine of the most widely spoken languages in the world. That pricing detail matters more than the model itself: it is the number that determines whether a back-office document pipeline is affordable at real volume.
Run the arithmetic on a corpus size enterprises actually have. A firm processing 50 million pages a year, the kind of volume a large insurer or bank generates from claims and statements, would pay roughly $75,000 annually through the Cohere API. Cohere’s own case study, a 13 million page monthly accounts payable workflow, claims Model Vault (its single-tenant deployment option) would save that customer $144,000 a year over API pricing, and $1.47 million a year against a hyperscaler charging $10 per 1,000 pages. Those figures are Cohere’s, not independently audited, but the order of magnitude explains why enterprises evaluate parsing costs per thousand pages rather than per document.
Document extraction is not an open field. Cohere is entering a category that already includes Mistral OCR, Databricks AI Parse, LlamaParse, and the hyperscaler offerings from AWS Textract and Google Document AI, plus open-weight options like Chandra OCR and DeepSeek-OCR. Cohere says Parse scored 79.2 on its own ParseBench evaluation, ahead of Mistral OCR 4 (74.5), Databricks AI Parse (72.4), and LlamaParse’s cost-effective tier (78.3), and more than 20 points ahead of both AWS Textract and Google Document AI.
Those benchmarks are Cohere’s own, scored on a rubric Cohere designed and revised. The company’s footnotes disclose that ParseBench’s scoring rules changed in August 2026 to fix a bug that had inflated Semantic Formatting results, and that Cohere re-scored every competitor’s outputs under the new rules rather than citing the vendors’ published numbers. That is a reasonable methodology, but it means every score in the comparison, including Cohere’s own, comes from a single company’s internal test.
The company is candid about where Parse loses. General-purpose frontier models, GPT-5.5, Opus 4.8, and Gemini 3.5 Flash, all outscored Parse on the same benchmark (84.4, 84.3, and 81.8 respectively). Cohere’s argument is that those models are larger and not purpose-built for high-throughput document pipelines, and that Parse’s real competition is specialized parsing tools, not frontier chatbots repurposed for OCR. On throughput, Cohere claims 4.5 pages per second per GPU, roughly 1.4 times the rate of RedNote’s dots.mocr and 2.2 times Chandra OCR 2 under the same hardware, though again the comparison is Cohere’s benchmarking setup.
Parse is available now through the Cohere API, Model Vault, Microsoft Foundry, and AWS SageMaker, with a free tier for testing through Cohere’s hosted Space. For teams building retrieval-augmented generation or agentic workflows that depend on clean document ingestion, per-page pricing this low removes cost as the primary objection to processing archival volumes rather than sampling them. Operators evaluating a parsing vendor for a six or seven figure page count should still request a proof-of-concept run against their own document mix rather than taking ParseBench at face value, since the benchmark and the vendor selling it are the same company.
Cohere disclosed pricing, benchmark results, and deployment options for Parse in a company blog post published on its own site.
