AI Rate CardHow much does RAG cost? Embeddings and reranking for document search, with live prices
Updated Nov 10, 2026
Retrieval-augmented generation (RAG) lets an assistant answer from your own documents: they are cut into passages, turned into vectors (embeddings), and at question time the closest passages are found, often reranked, and given to the answer model. The RAG cost calculator prices the embedding and reranking part at every embedding and rerank model, at the cheapest seller of each.
How to use it
- Pages and tokens per page: a page of plain text is about 500-700 tokens.
- Changed a month: the share of documents re-embedded as they are edited or added.
- Queries a day and tokens per query: every question is embedded too.
- Passages reranked per query: 0 if you do not rerank; 10-50 is common, each passage a few hundred tokens.
Example from today's prices
10,000 pages (6M tokens), 10% changed a month, 1,000 queries a day, 20 passages reranked:
- Embedding the documents once: from $0.03 (small open models such as BGE or Qwen3 Embedding 0.6B), $0.06 for Qwen3 Embedding 8B, about $0.10 for OpenAI text-embedding-3-small, $0.62 for text-embedding-3-large, $0.72 for Cohere Embed v4, about $0.90 for Gemini Embedding.
- Every month (changes + queries, 1.5M tokens): cents.
- Reranking (258M tokens a month): about $2.60 with Qwen3 Reranker 0.6B or BGE Reranker, about $13 with Cohere Rerank 5 Pro or Jina Reranker v3. Embeddings are rarely the expensive part. The bill is the reranker and above all the answer model that reads the passages - price it with the coding agent calculator or the cost calculator.
Choosing a model
- Quality first: test two or three embedding models on your own questions - a cheap open model is often good enough for one language.
- Many languages: BGE M3, Qwen3 Embedding, Cohere Embed v4, Gemini Embedding.
- Keep the same model: changing the embedding model means re-embedding everything - the "index once" cost again.