Embed 5 Pro

Not yet independently verified. Every detail here comes from Cohere’s own announcement and documentation; no independent report or measurement of Embed 5 was found. Its benchmark results are Cohere’s own. We will update this when it can be confirmed, and remove this note.

Embed 5 Pro is the larger of two embedding models Cohere released on 30 September 2026 for enterprise search, retrieval-augmented generation and classification. According to Cohere’s announcement, it embeds text, images, fused text and images, and parsed documents, in more than 100 languages with particular attention to European languages and wider coverage of Japanese, Chinese, Korean, Arabic, Farsi, Hindi, Bengali, Telugu, Indonesian and Thai. It accepts up to 128,000 tokens of context, returns embeddings of 2048, 1536, 1024, 768, 512 or 256 dimensions, and can output float, int8 or binary vectors, so storage can be traded against accuracy. It is available on Cohere’s API and Model Vault, Microsoft Foundry, Amazon SageMaker and Cohere’s North platform, and for private deployment, with integrations for the main vector databases and search engines. Cohere prices it at $0.12 per million text tokens and $0.40 per million image tokens. Cohere reports leading results on visual-document, financial and parsed-PDF retrieval evaluations; these are its own figures and are not stored as scores. The board also lists Cohere’s earlier Embed v3.