Cohere releases Parse 5 with 79.2 ParseBench score

• Parse 5 is a 2.3B-parameter vision-language model with an 8,192-token context window and an approximately 4.6 GB footprint, according to Cohere's announcement.
• Cohere reports a 79.2 ParseBench score across three dimensions, ahead of Mistral OCR 4 at 74.5 and Databricks AI Parse at 72.4, in its benchmark post.
• The API costs $1.50 per 1,000 pages, with Cohere's pricing post claiming up to 95% lower pricing than frontier LLMs and hyperscaler offerings.
• Parse returns reading-order Markdown, HTML tables, image descriptions, and pixel or normalized bounding boxes, Cohere's capabilities post says.
• It is generally available through the Cohere API, Model Vault, AWS SageMaker, and Microsoft Foundry, according to Cohere's availability post.

ParseBench's public site measures semantic correctness for agent workflows, not only text similarity. The API reference exposes a blocks mode with typed regions and coordinates, while Cohere's launch post says Model Vault can cut inference costs by 61% at full hourly utilization. The migration gotcha is in the docs: the API currently accepts image_url inputs, even though the model documentation lists PDF and PPT support.

Parse 5 is aimed at enterprise ingestion, search, retrieval-augmented generation, and document agents rather than general-purpose text generation. Cohere's launch post describes it as a high-throughput vision parser for finance, insurance, scientific work, and other business documents.

Cohere's model documentation identifies the shipped model as parse-v5.0 and lists:

• A proprietary north-micro-vision-instruct architecture.
• 2.3 billion parameters and an approximately 4.6 GB model size.
• An 8,192-token context length.
• Stable parsing for Arabic, English, French, German, Japanese, Korean, Italian, Portuguese, and Spanish.
• Zero-shot support for additional languages, with potentially lower accuracy.

The model docs describe Parse as a companion to Cohere's Embed and Rerank models for search pipelines and Compass deployments.

One post aidangomez's post describes parsing as the bottleneck between enterprise data and effective LLM use, with a familiar tradeoff between expensive high-quality understanding and scalable lossy extraction. Parse's response design targets that handoff directly.

The API reference defines two response shapes:

• markdown: Pages arrive in document order. Text follows reading order, tables are inlined as HTML, and images are referenced with IDs plus descriptions and metadata.
• blocks: Each page contains an ordered array of text, image, and table regions. Image and table blocks include pixel-coordinate and normalized bounding boxes; tables are serialized as HTML.
• Page metadata: The response includes page indexes, API version information, token telemetry, and billed page counts.

The output is designed for rendering and downstream parsing. Cohere's documentation says arbitrary structured JSON output and confidence scores are not supported.

ParseBench evaluates document parsers on roughly 2,000 human-verified enterprise pages and more than 169,000 test rules across five dimensions: tables, charts, content faithfulness, semantic formatting, and visual grounding, according to the benchmark's GitHub repository.

Cohere's published 79.2 figure is a three-dimension average, covering tables, content faithfulness, and semantic formatting. Its table reports:

• Tables: 87.0.
• Content faithfulness: 86.6, covering omissions, hallucinations, and broken reading order.
• Semantic formatting: 64.0, covering styles such as strike-throughs and italics when they change meaning.

The same comparison puts Parse above several specialized systems:

• Mistral OCR 4: 74.5, a 4.7-point gap.
• Databricks AI Parse: 72.4, a 6.8-point gap.
• Azure Document Intelligence: 69.3, a 9.9-point gap.
• LlamaParse Cost Effective: 78.3, a 0.9-point gap.

The frontier models in Cohere's table score higher on this narrower comparison: GPT-5.5 at 84.4, Opus 4.8 at 84.3, and Gemini 3.5 Flash at 81.8. Cohere says it re-scored competitors after an August 2026 ParseBench rule fix for bold and heading detection, so 79.2 should be read as Cohere's updated three-dimension comparison, not the full five-dimension benchmark score. The launch post also says the comparison excludes chart extraction and the spatial layout or visual-grounding dimension because those capabilities sit outside Parse's current product scope.

Cohere prices API access at $1.50 per 1,000 pages. Its launch post says that is up to 95% below frontier LLM and hyperscaler offerings and 63% below Mistral, all figures reported by Cohere.

The same launch post reports 4.5 pages per second in its throughput test, or 36 pages per second on an eight-H100 node. It says Parse was about 1.4 times faster than dots.mocr and 2.2 times faster than Chandra OCR 2 on the tested hardware, with open-source models served through vLLM.

Model Vault uses dedicated instances rather than per-page API billing. Cohere's pricing page lists Parse 5 tiers at $4 per hour or $2,500 per month, and $7 per hour or $4,300 per month.

Cohere reports that Model Vault costs 23% less than the API at 50% GPU utilization and 61% less at full hourly utilization. Its example of a 13-million-page-per-month accounts-payable workflow comes out to approximately $12,000 in monthly savings against the API, or $1.47 million annually against a hyperscaler priced at $10 per 1,000 pages.

The product documentation and API reference expose an important distinction between the model's file support and the current endpoint contract.

• The model overview lists PDF, PPT, and JPEG as supported file types, encoded as Base64.
• The API reference currently accepts only document.type = image_url, using either a remote HTTP(S) image URL or a data URI. It explicitly says PDF and file URL inputs are not yet supported.
• The endpoint limits each image to 20 MB and 50 megapixels or 200 MB decoded, whichever limit is exceeded first.
• The quickstart demonstrates a Base64-encoded PNG sent to /v2/parse, rather than a PDF upload.
• Cohere's documentation says Parse does not identify headers, footers, or font hierarchy, and returns Markdown rather than user-defined structured JSON.

That leaves document-level ingestion and page-image parsing as separate integration layers in the current documentation. Compass advertises broader formats, while the public Parse endpoint documents an image-only request shape.

Parse 5 is generally available through:

• The Cohere API.
• Model Vault for managed, single-tenant inference.
• AWS SageMaker.
• Microsoft Foundry.
• A free Hugging Face Space.
• Cohere's North and Compass products.

The Compass documentation places Parse alongside Embed and Rerank in a managed search stack that covers document ingestion, parsing, chunking, indexing, hybrid search, and two-stage retrieval. Cohere also invited builders to try the model and report back in its support post.