Embed 5 Pro

Cohere released Embed 5 on October 1, 2026, a new embedding family with Pro and Fast tiers aimed at enterprise retrieval. Pro is optimized for maximum quality across multimodal, multilingual, financial, code, and parsed-document retrieval, while Fast targets latency- and cost-sensitive workloads. Both tiers share a single embedding space, enabling teams to index with Pro and query with Fast without rebuilding the index. Embed 5 Pro delivers Cohere's strongest retrieval performance to date, achieving the highest average score across ViDoRe V3, financial documents, parsed PDFs, image retrieval, and key business languages in internal testing. It is the first model family evaluated with Cohere's new RCP-nDCG@10 retrieval methodology, which scores documents against query-specific relevance criteria rather than fixed labels. The model is generally available on the Cohere API and Model Vault, Microsoft Foundry, and Amazon SageMaker at $0.12 per million tokens.

Embed 5 Pro carries the API model ID embed-v5.0-pro and replaces Embed 4 as Cohere's flagship embedding model. It supports text, images, and fused text+image inputs with a 128K token context length and output dimensions ranging from 256 to 2048 (default 2048) in float, int8, or binary formats. The model covers 100+ languages and is priced at $0.12 per million text tokens and $0.40 per million image tokens. Unlike its predecessor Embed 4, which maxed at 1536 output dimensions and 47 text languages, Embed 5 Pro offers expanded dimensional flexibility and broader language support while maintaining the same 128K context window. This makes Pro suitable for enterprise RAG pipelines that need to index large multilingual corpora with high-fidelity vector representations. The spec parity with Fast on context and modalities lets teams swap between tiers for cost or latency reasons.

Embed 5's Pro and Fast variants share the same vector space, letting teams index with Pro and serve live queries with Fast without switching between incompatible embedding models. Cohere recommends this split to favor retrieval quality during preparation and low latency during interactive search. This design avoids the costly re-indexing migration that typically locks teams into older embedding models. The family handles text, images, and mixed inputs including PDF pages, with 100+ language support and a 128,000-token input window. It improves on Embed 4 for visually rich documents, financial filings, parsed PDFs, code, and multilingual retrieval. The models are available through the Cohere Embed API, Microsoft Foundry, and Amazon SageMaker, with Model Vault for single-tenant deployment, though Cohere did not publish a quantitative Pro-versus-Fast tradeoff figure.

On the fused text-image average benchmark, Cohere Embed 5 Pro leads with an nDCG@10 score of 82.3, followed by Embed 5 Fast at 81.2 and Gemini Embedding 2 at 61.3. These figures come from Cohere's provider report and were captured on the October 1, 2026 launch date. The benchmark evaluates text queries retrieving embeddings that combine page images with metadata across five datasets. Pro's lead on fused multimodal retrieval is relatively narrow over Fast, suggesting the two tiers trade off modestly on this specific metric despite their shared embedding space. The evaluation snapshot, hardware, and dates are undisclosed, limiting independent verification. The BenchLM page also includes a RepairBench multilingual subset and a High Finance internal dataset in the fused average calculation.

Cohere's Embed 5 Pro achieves a reported ViDoRe V3 average score of 85.8, compared to 84.5 for the Fast variant, on enterprise document retrieval benchmarks. The model is designed for quality-focused retrieval across multimodal, multilingual, financial, code, and parsed-document workloads. Its shared embedding space with Fast allows enterprises to index with Pro and query with Fast without vector incompatibility. Deployment options for Embed 5 Pro include the Cohere API, Model Vault, Microsoft Foundry, Amazon SageMaker, and private infrastructure. The two-tier architecture lets engineering teams separate expensive offline indexing from high-volume online querying, which is critical for agentic workflows where retrieval cost and speed compound across repeated search steps. Cohere positions Embed 5 as a retrieval foundation for search, RAG, and agentic applications.