Cohere Parse 5 turns documents into Markdown at $1.50/1K pages
A 2.3B vision model does single-pass OCR, tables and layout for RAG far below GPT-5.5's cost — plus Z.ai's MIT-licensed multimodal GLM-5.3-Flash open weights.
Copy markdown
$1.50 per 1,000 pages, in a single pass
Parse 5 (parse-v5.0) is a 2.3B vision-language model that turns PDF, slide and JPEG pages into Markdown — with HTML-rendered tables, form key-values, image descriptions and bounding boxes — in one pass, with no separate OCR stage. At $0.0015 a page it's priced for high-volume document ingestion, not one-off parsing.
Loses the benchmark, wins on cost per page
It scores 79.2 on ParseBench, behind GPT-5.5 (84.4), Opus 4.8 (84.3) and Gemini 3.5 Flash (81.8) but ahead of Mistral OCR 4 (74.5) and Azure Document Intelligence (74.3). Cohere pegs one large financial-docs pipeline at ~98% cheaper than routing pages through GPT-5.5 (its own estimate) — for RAG at scale, that accuracy gap is often worth trading.
GA now — API, SageMaker, or self-hosted
No waitlist: it's live on the Cohere Parse API, AWS SageMaker, Microsoft Foundry and a single-tenant Model Vault, with a free trial key and a Hugging Face Space to test. The model is ~4.6GB with an 8K context and 9 stable languages. Weekend build: point it at a folder of scanned PDFs and pipe citation-traceable Markdown straight into your vector store.
Elsewhere: GLM-5.3-Flash lands as MIT multimodal weights
Z.ai's new Flash variant — 320B total, ~18B active, 1M context — is the first natively multimodal model in the GLM-5 series, and it ships under a permissive MIT license, unlike last week's more restrictively-licensed GLM-5.3 base. transformers v5.16.1 already supports it.