Embeddings Cost Calculator

Embedding the corpus is the cheap part

The first instinct is to fear the cost of embedding a big knowledge base. Relax: modern embedding models are extraordinarily cheap per token — often two to thirteen cents per million tokens. Embedding 100,000 chunks of 500 tokens is 50 million tokens, which lands somewhere between $1 and $7 as a one-time cost on most providers. For typical corpora the one-time embed is a rounding error. The costs that actually add up live somewhere else, and they're easy to miss when you're estimating.

The two costs people forget

First, every query gets embedded too. A search product embedding 500,000 queries a month is paying a small recurring fee that, over a year, can quietly exceed the one-time corpus cost. Second, changing your embedding model or chunking re-embeds everything — switch from a small model to a large one, or re-chunk for better retrieval, and you pay the full corpus cost again. Teams iterating on retrieval quality re-embed several times before they're happy. Set your expected re-embeds per year above to see what that iteration really costs.

Storage is the bill that never stops

Embeddings are vectors, and storing millions of them in a managed vector database — Pinecone, Qdrant Cloud, Weaviate — is usually the largest ongoing line, dwarfing the embedding API itself. Higher-dimension models (1536, 3072) cost more to store and search than compact ones, so the cheapest embedding model per token isn't automatically the cheapest system. Price the storage and query side on the vector database cost calculator, then put the whole pipeline together on the RAG cost calculator.

Related tools & guides

Vector database cost calculator · RAG cost calculator · Fine-tuning cost calculator · LLM token cost calculator · All AI APIs

Frequently asked questions

How much does it cost to embed documents?

Embedding is billed per token, like an LLM input. Total tokens = documents × average tokens per document, then × the provider's price per million tokens. Modern embedding models are cheap — often $0.02 to $0.13 per million tokens — so embedding a few million tokens once costs cents to low dollars. The cost only matters at very large corpora or when you re-embed often.

What's the hidden recurring cost of embeddings?

Two things people forget. First, every search query must also be embedded — at high query volume that recurring cost can exceed the one-time corpus embedding. Second, whenever you change embedding model or chunking strategy you must re-embed the entire corpus from scratch, paying the full cost again. Vector database storage is separate and is usually the larger ongoing bill.

Which embedding model is cheapest?

Among hosted APIs, small models like OpenAI text-embedding-3-small and Google's embedding model are among the cheapest per token, while larger models like text-embedding-3-large and Voyage's flagship cost more but capture more nuance. Cheapest per token isn't always cheapest overall — a higher-quality embedding can need fewer chunks and less re-ranking. Compare price against your retrieval quality, not in isolation.

One-time corpus cost by provider