Cohere has launched the 2.3 billion-parameter Parse 5 model for converting PDFs, slides and scanned documents into structured Markdown. The model processes the page as an image and produces results in a single vision-language model pass; it returns tables as HTML, images with descriptions and coordinates indicating the locations of tables and images. The API price was announced as $1.50 per 1,000 pages.
In Cohere’s ParseBench comparison, Parse 5 scored 79.2 across tables, content accuracy and semantic formatting. This score ranked behind GPT-5.5’s 84.4, Opus 4.8’s 84.3 and Gemini 3.5 Flash’s 81.8, but ahead of LlamaParse’s 78.3 in the Cost Effective tier. The model is available through the Cohere API, Model Vault, Microsoft Foundry and AWS SageMaker. While the Chart and Layout dimensions were not included in the evaluation, it was stated that chart data extraction is planned for a future version.
Why it matters
Parse 5 offers an output standard that enables tables, images and their positions on the page to be processed together, rather than merely extracting text in document-conversion workflows. This could directly affect how the resulting data is used in subsequent processes by teams working with PDFs and scanned documents; however, the lack of chart data extraction support currently limits its scope. The performance results show that the model ranks lower than some of its competitors but higher than LlamaParse’s specified tier; in addition, the absence of chart and layout dimensions from the evaluation narrows the scope of the comparison. Pricing and access across different cloud platforms mean that the choice must be based not only on the model score but also on the way it will be used and the existing technical infrastructure.