Parse is built for teams processing hundreds of thousands of document pages a month on an 8,192-token context budget per page, not for one-off uploads or long-document reasoning. It replaces dedicated OCR tools like AWS Textract and Google Document AI in RAG and document pipelines where cost and throughput matter more than top-end accuracy.
Parse is Cohere's 79.2-ParseBench vision-language model that converts PDF, PPT and JPEG document pages into structured Markdown with HTML tables, form key-value pairs and bounding boxes. Built for high-volume document ingestion rather than general chat, it outperforms Mistral OCR 4 and AWS Textract while trailing frontier multimodal LLMs on raw accuracy.
Provider: Cohere · Family: Parse
Context window: 8,192 tokens
Input modalities: pdf, image · Output: text
About Cohere Parse 5
Cohere released Parse (model ID parse-v5.0) on August 27, 2026: a 2.3-billion-parameter vision-language model built on Cohere Labs' proprietary North-Micro-Vision-Instruct architecture, with a roughly 4.6GB footprint, purpose-built to turn document pages into Markdown rather than to chat. It sits outside Cohere's Command lineup entirely, one job only, page-to-structured-text conversion, meant as the ingestion step ahead of a RAG, search or agent stack. On Cohere's own ParseBench, Parse scores 79.2 overall (87.0 tables, 86.6 content faithfulness, 64.0 semantic formatting), ahead of Mistral OCR 4 (74.5), Databricks AI Parse (72.4), LlamaParse (78.3), Google Document AI (57.3) and AWS Textract (53.3), but behind general-purpose multimodal LLMs sometimes used for parsing, GPT-5.5 (84.4) and Gemini 3.5 Flash (81.8), which cost far more per page. Parse works page by page: its context window is 8,192 tokens, enough for one dense page of text and layout metadata, with no long-context recall to benchmark since the model transcribes and structures rather than answers questions about a document. It accepts PDF, PPT and JPEG pages submitted as base64 data, reads text in reading order, renders tables as HTML, extracts lists and form key-value pairs, describes embedded images, and returns bounding boxes for tables and images so a downstream app can cite the exact source region. It covers nine languages at full accuracy, including Arabic and Korean, with weaker zero-shot support elsewhere, and does not detect headers, footers, font-based heading hierarchy, or report per-field confidence scores. Billing is by the page rather than by token, and Model Vault offers a cheaper single-tenant deployment once volume climbs; anyone can try the model for free first on Cohere Labs' Hugging Face Space (see the pricing FAQ for exact figures). Parse shipped generally available across four paths on day one, the Cohere API, Model Vault, Microsoft Foundry on Azure and Amazon SageMaker, with no waitlist or research license, and throughput reaching 4.5 pages per second on a single instance and up to 36 pages per second on an 8xH100 node. Cohere's eight SDKs, Python, TypeScript, Go, Java, Ruby, PHP, C# and Swift, expose the endpoint via a standard bearer key, capped at 20MB or 50 megapixels per file. Cohere carries a company-wide SOC 2 Type II report, ISO 27001 and ISO 42001 certification, and a GDPR data processing addendum on request. By default, prompts sent through the shared API may be used to improve Cohere's models unless a customer opts out via dashboard controls; zero data retention is available on approval, and Model Vault or an on-premises deployment keeps documents inside the customer's own environment entirely. Parse's training-data mix isn't detailed in any public system card the way Cohere documents its Command chat models; what Cohere has shared is architectural and benchmark-based rather than a data-lineage disclosure. Parse fits RAG, search and document-heavy agent teams processing hundreds of thousands of pages a month in finance, insurance, healthcare, public sector and manufacturing, where a predictable per-page price and page-level bounding boxes matter more than squeezing out the last points of accuracy. It is the wrong pick for teams that need automated confidence-based quality gating, the highest possible accuracy on messy scans regardless of cost, or a single end-to-end platform, since Parse performs only the extraction step and leaves embedding, retrieval and generation to whatever the team builds around it. It launched as a single generally available version with no predecessor publicly documented and no successor announced as of this writing.
Pricing
$1.50 per 1,000 pages on the pay-as-you-go API, since Parse bills by page instead of by token. Model Vault trades a flat monthly fee, $2,500 for Medium or $4,300 for XL, for lower per-page cost once usage passes about 1.67 to 2.87 million pages a month. A free CohereLabs Hugging Face Space covers testing before committing to either paid path.
Key Features
- HTML Table Extraction: Renders every detected table as HTML embedded directly in the Markdown output, scoring 87.0 on ParseBench's table-extraction sub-benchmark.
- Bounding-Box Grounding: Returns pixel coordinates for tables and images so downstream apps can highlight or cite the exact source region for a fact.
- Form Key-Value Extraction: Pulls structured key-value pairs out of forms and invoices without a separate template or training step per document type.
- Nine-Language Coverage: Delivers full accuracy across nine languages, among them English, Spanish, Japanese and Arabic, with weaker zero-shot support outside that set.
- High-Throughput Deployment: Hits 36 pages per second when running on Cohere's largest Model Vault tier, suited to ingesting hundreds of thousands of pages a month.
Pros
- Costs a small fraction of what general-purpose multimodal LLMs charge for the same parsing job, per Cohere's published per-page rate.
- Beats dedicated OCR competitors, including Mistral OCR 4 and Databricks AI Parse, on Cohere's own ParseBench table and content-fidelity scores.
- Available day one across four deployment paths, API, Model Vault, Microsoft Foundry and AWS SageMaker, with no waitlist or research license.
Cons
- No per-field confidence scores, so regulated pipelines need their own quality-gating layer on top.
- Trails frontier multimodal LLMs on raw ParseBench accuracy, so pixel-perfect extraction on hard documents may need a heavier, pricier model instead.
- Does not detect headers, footers or font-based heading hierarchy, so page furniture can end up inline with body text.
Benchmarks
- parsebench tables: 87
- gpt 5 5 parsebench: 84.4
- parsebench overall: 79.2
- llamaparse parsebench: 78.3
- aws textract parsebench: 53.3
- mistral ocr 4 parsebench: 74.5
- gemini 3 5 flash parsebench: 81.8
- google document ai parsebench: 57.3
- databricks ai parse parsebench: 72.4
- parsebench semantic formatting: 64
- parsebench content faithfulness: 86.6
- throughput pages per sec 8x h100: 36
- throughput pages per sec single instance: 4.5
Frequently Asked Questions
What are Cohere Parse 5's pricing plans in 2026?
Cohere bills the pay-as-you-go API at $1.50 for every 1,000 pages parsed, charging by page rather than by token. For steadier high-volume use, Model Vault (a single-tenant dedicated deployment) costs $2,500 monthly on the Medium tier or $4,300 monthly on the XL tier, undercutting the API past roughly 1.67 to 2.87 million pages each month. Cohere has not published any additional volume discount or annual-contract pricing beyond those two Model Vault tiers.
Is Cohere Parse 5 free to use?
There is no free quota on the production API or Model Vault; every page processed through those paths is billed. Cohere does offer a free Hugging Face Space where anyone can upload a document and see Parse's output before paying for API access, useful for evaluating quality but not for production volume.
What are Cohere Parse 5's closest competitors?
Parse's closest rivals are dedicated document-parsing models: Mistral OCR 4, Databricks AI Parse and LlamaParse all target the same RAG-ingestion use case at broadly similar prices. Teams that need higher raw accuracy and do not mind a much higher per-page cost sometimes route parsing through general-purpose multimodal LLMs like GPT-5.5 or Gemini 3.5 Flash instead. Legacy OCR services such as AWS Textract and Google Document AI remain cheaper for plain text but handle tables and layout far less reliably.
Cohere Parse 5 or Mistral OCR 4: which should you pick?
Parse scores higher than Mistral OCR 4 on Cohere's own ParseBench, particularly on table extraction, and comes in cheaper on Cohere's published per-page rate. Mistral OCR 4 counters with much broader language coverage and its own self-hosted deployment path, so teams already standardized on Mistral's stack have less reason to switch just for Parse's table-extraction edge.
How long does it take to get going with Cohere Parse 5?
Sign up for a Cohere API key from the dashboard, then send a page as a base64-encoded image or PDF to the parse endpoint using any of Cohere's eight SDKs or a plain HTTP call, specifying the markdown or the less-documented blocks output format. There is no waitlist or research license to clear, so a first successful parse typically takes under an hour once you have a key. Teams wanting to test quality first can instead upload a sample document to the free CohereLabs Hugging Face Space.
Top Alternatives
- Mistral OCR 4: Pick Parse if you want a higher ParseBench table score and a lower per-page price; pick Mistral OCR 4 if you need broader language coverage or self-hosting.
- GPT-5.5: Pick Parse for high-volume ingestion at a fraction of the cost; pick GPT-5.5 when you need the highest possible parsing accuracy and can absorb the price gap.
- Gemini 3.5 Flash: Pick Parse when page volume and cost predictability matter most; pick Gemini 3.5 Flash if you also need general multimodal reasoning beyond document parsing.