01
What Cohere Parse 5 is
Cohere introduced Parse on August 27, 2026 as a specialist vision-language model for document ingestion. The current model ID is parse-v5.0. Cohere documents a 2.3B-parameter model, an 8,192-token context window and an approximate 4.6GB model size.
The product focuses on preserving reading order and document structure for search, RAG and agent workflows. It extracts text, tables, lists, forms, key-value pairs, images and captions, with Markdown as the default output and an optional ordered blocks format.
02
API availability and the current input contract
Cohere lists Parse availability through its API, Microsoft Foundry, AWS SageMaker and Model Vault. The public endpoint is /v2/parse. Cohere's current rate-limit table lists Parse at 500 requests per minute for both trial and production API keys.
There is an important documentation nuance: the product overview lists PDF, PPT and JPEG support through Base64-encoded document representations, while the live API reference currently says the request accepts document.type = image_url. Developers should follow the live endpoint contract for implementation rather than assume arbitrary PDF or file URLs can be submitted directly.
03
Cohere Parse pricing
Cohere's August 27 launch post states that the public Parse API costs $1.50 per 1,000 pages, or $0.0015 per page before storage, retrieval, embedding, reranking and downstream generation costs.
Cohere also promotes Model Vault for single-tenant inference and says it can reduce effective cost at sustained utilization, but the launch article does not give one universal Model Vault price. Deployment-specific commercial terms should be obtained rather than extrapolated from the public API rate.
04
Benchmarks and limits
Cohere reports a 79.2 average across the ParseBench dimensions it treats as aligned with the product: tables, content faithfulness and semantic formatting. Those results, as well as Cohere's throughput claims, are vendor-published and should not be presented as an independent Project Monet benchmark.
Cohere says Parse does not provide extraction confidence scores and does not target chart-to-table numerical extraction in the current product scope. High-stakes amounts, dates, account numbers and legal clauses therefore still need deterministic or human validation.
05
Where Parse fits in an AI stack
A common workflow is document input → Parse → structured chunks → embeddings → retrieval → reranking → answering model or agent. Parse handles document understanding; it does not replace the retrieval or reasoning stages that come afterward.
For an implementation walkthrough covering Python, data-URI input, Markdown versus blocks and production safeguards, read the Cohere Parse API guide. Cohere's public material does not establish that parse-v5.0 is an unrestricted open-weight download, so private deployment should not be confused with open weights.
Sources
Primary and supporting sources
Facts were rechecked against the linked sources immediately before publication. Pricing, product availability and rollout status can change.