Cohere has released Embed 5, its latest embedding model family, split into two versions aimed at different stages of a retrieval pipeline. Embed 5 Pro targets indexing and quality-critical search, while Embed 5 Fast is built for low-latency, high-volume queries. Both produce vectors in the same embedding space, so a team can index a document collection with Pro and then query it with the cheaper Fast model, without rebuilding the index from scratch.
Cohere says that pairing costs around 1.6 points of retrieval quality compared with using Pro throughout, by its own internal measurements. Both models handle text, images and mixed text-and-image inputs, support more than 100 languages, offer a 128,000-token context window, and let developers choose output dimensions between 256 and 2,048 depending on storage and accuracy needs.
Pricing is set at $0.12 per million text tokens for Pro and $0.08 for Fast, with image inputs priced at $0.40 per million tokens on either model. Embed 5 is available through Cohere's own Embed API, through Microsoft Foundry and Amazon SageMaker, and as a single-tenant deployment via Cohere's Model Vault for customers who need isolated infrastructure.
On Cohere's own fused text-and-image benchmark, Pro scored 82.3 and Fast 81.2, ahead of the 61.3 the company reports for Google's Gemini Embedding 2. As with any vendor-run benchmark, those figures are worth treating as a starting point rather than a final word.
Source: MarkTechPost
