Cohere Embed 5 Launch: Pricing & RAG Impact

Cohere launched Embed 5 Pro and Embed 5 Fast on October 8, 2026. This is a retrieval-model release, not a new chat model. The immediate buyer decision is whether better search quality and a mixed Pro/Fast architecture justify re-embedding an existing corpus.

The live pricing table above reads our maintained dataset at build time. Check Cohere pricing for the full model catalog and current rates; use the token calculator to size the text portion of an indexing workload. Image-heavy documents and managed-cloud contracts need separate estimates.

Route	Best starting role	Main cost to include
Embed 5 Pro	Quality-critical corpus indexing	Initial and incremental reindexing
Embed 5 Fast	Interactive query embedding	Request volume and latency
Embed v3	Existing production baseline	Migration and vector compatibility

What changed

Cohere’s release notes name two API model IDs: embed-v5.0-pro and embed-v5.0-fast. Pro is tuned for highest retrieval quality, particularly offline indexing; Fast is tuned for latency and throughput in interactive search and agent loops. Both map into a shared embedding space. Cohere recommends indexing documents with Pro and embedding queries with Fast, so a team can avoid paying the Pro route for every query.

The family accepts text, images, and mixed text-and-image inputs, including PDF pages. Cohere says it supports more than 100 languages, a 128,000-token context window, and Matryoshka output dimensions from 256 through 2,048, with float, int8, and binary formats. Those are vendor specifications, not a guarantee that any chosen dimension preserves your application’s recall.

The release notes claim gains over Embed 4 on visually rich documents, financial filings, parsed PDFs, code, and multilingual retrieval. Cohere did not provide an application-specific saving or a universal quality lift; test those claims on your own corpus.

Pricing impact: indexing versus querying

There is no useful single “RAG price.” The table above shows tracked embedding API rates, but a migration can add a one-time reindexing bill, vector-storage writes, dual-index capacity during cutover, reranking, answer-generation tokens, and engineering review. If the new model improves top-k recall enough to shrink rerank depth or failed-answer retries, those downstream savings can outweigh its embedding rate. If retrieval quality barely changes, a large reindex may not pay back.

Keep Pro and Fast costs separate in your ledger. Record indexed tokens or documents, query volume, storage size at the selected dimension, p95 latency, recall at k, and cost per accepted answer. Compare the mixed Pro-index/Fast-query setup with all-Fast and your current model under identical evaluation data.

Who should switch now?

New enterprise search projects can evaluate Embed 5 directly, especially for multilingual or visually rich documents. Existing Embed users should run a shadow index before replacing production vectors; changing embedding families without rebuilding the relevant index can break similarity quality. High-volume query systems should test whether Fast maintains recall against a Pro-built corpus while meeting latency targets.

Embed 5 is available through Cohere’s Embed API, Microsoft Foundry, Amazon SageMaker, and Cohere Model Vault for single-tenant deployment. Prices and commitments may differ by route, so do not treat the direct API table as a cloud-marketplace quote.

For alternatives and full-stack tradeoffs, see our Cohere API pricing guide, OpenAI pricing, and best AI for RAG guide.

Bottom line

Embed 5 creates a practical two-tier retrieval option: index for quality, query for speed. Migrate only after a representative recall and cost-per-answer test shows a gain over your current stack.

Source checked October 8, 2026: Cohere’s official Embed 5 release notes. Model IDs and tracked API rates are present in our live pricing dataset.