How are embedding models priced?

Per input token only — there is no output token charge, which makes embeddings far cheaper per call than chat completions.

What drives embedding cost in a RAG system?

Corpus size at index time, and query volume plus chunk size at retrieval time. Re-indexing a large corpus is usually the single biggest line item.

Is a larger embedding model worth it?

Only if retrieval quality measurably improves for your corpus. Price both dimensions — model rate and storage or index cost — before upgrading.

Embedding Models Pricing: Per-Token Rates Across Providers

Current per-token rates

OpenAI text-embedding-3-small: $0.02/1M. text-embedding-3-large: $0.13/1M. Cohere embed-v3: $0.10/1M. Voyage 3-large: $0.18/1M. Gemini text-embedding-004: $0.025/1M.

Cost per million vectors

At an average chunk of 500 tokens, 1M vectors ≈ 500M tokens. With OpenAI small, that's ~$10. With Voyage large, ~$90. Pick by recall need, not vibes.

Re-embedding cost

Most teams forget that schema changes, chunking changes, and model upgrades require full re-embed. Budget 2–4× initial embedding cost over a year.

Storage and query cost

Vector DB (Pinecone, Weaviate, pgvector) cost often exceeds embedding cost. A 10M-vector index runs $200–$800/mo across hosted providers.

Picking the right embedder

For most RAG: OpenAI 3-small or Gemini 004 — cheapest with 95th-percentile recall. For high-stakes or domain-specific: Voyage 3-large.

Worked example: re-embedding a growing corpus

A corpus of 2M chunks at 500 tokens each is 1B tokens. Initial embedding with OpenAI text-embedding-3-small ($0.02/1M) costs $20. If chunking strategy changes twice a year and the model is upgraded once, that's three full re-embeds in a year: 3*$20=$60 — trivial at this rate, but the same math on Voyage 3-large ($0.18/1M) is 3*$180=$540, a meaningfully different annual line item for the same corpus.

Mechanics: why embedding cost scales differently from generation cost

Embedding models have no output tokens — the entire cost is the input pass that produces a fixed-length vector, so there's no 3-5x output multiplier the way there is with generative models. This is why embedding rates look tiny per token (fractions of a cent per 1M) but still accumulate meaningfully when a pipeline re-embeds the same corpus repeatedly rather than once.

Decision checklist before picking an embedding model

— Does your chunking strategy change often enough that re-embed cost matters? — Have you compared cost per million vectors, not just cost per token, across candidates? — Does the use case need Voyage/Cohere-tier recall, or does a $0.02/1M model clear your accuracy bar? — Have you separately budgeted vector database storage and query cost, which often exceeds the embedding cost itself?

What this post does not cover

This covers per-token embedding rates and re-embed cost mechanics; it does not cover vector database pricing structures in depth (those vary by provider and index type) or retrieval quality benchmarks, which determine whether a cheaper embedding model is actually a false economy for a given use case.

Frequently asked questions

How are embedding models priced?: Per input token only — there is no output token charge, which makes embeddings far cheaper per call than chat completions.
What drives embedding cost in a RAG system?: Corpus size at index time, and query volume plus chunk size at retrieval time. Re-indexing a large corpus is usually the single biggest line item.
Is a larger embedding model worth it?: Only if retrieval quality measurably improves for your corpus. Price both dimensions — model rate and storage or index cost — before upgrading.