What it costs to embed a corpus
Indexing cost is the corpus size in tokens times the per-token rate. The table uses standard rates; providers with a Batch rate cost less if indexing can wait.
As a rough guide, 1M tokens is about 750,000 English words, or around 1,500 pages of prose. Code, tables and non-English text usually need more tokens per word; count a sample with the Token Calculator.
How to compare embedding models on cost
- Price the one-time index separately from the monthly cost of new documents and search queries.
- Use Batch for indexing when the provider offers it; queries need real-time embeddings at the standard rate.
- Larger vectors cost more to store and search in your vector database, which often outweighs the embedding price itself.
- Switching models later means re-embedding the whole corpus, so include one full re-index in your plan.
Related pricing pages
Embeddings are usually the small line in a RAG budget. To price the generation side, compare chat models in the cheapest LLM API comparison or estimate a full workload with the AI API Cost Calculator. Provider pages list every model with its sources: OpenAI, Google Gemini, Mistral AI and Cohere. The raw records are in the AI API Pricing Dataset.
Frequently asked questions
Which embedding API is the cheapest?
OpenAI text-embedding-3-small has the lowest published text rate in this comparison at $0.02 per 1M input tokens. The most expensive text rate here is Google Gemini Embedding 2 at $0.2. Retrieval quality differs between models, so test a sample of your own data before choosing on price.
How much does it cost to embed 1 billion tokens?
At standard rates, 1 billion tokens costs $20.00 with text-embedding-3-small and $200 with Gemini Embedding 2. One billion tokens is roughly 750 million English words.
What is the difference between text-embedding-3-small and text-embedding-3-large pricing?
OpenAI charges $0.02 per 1M tokens for text-embedding-3-small and $0.13 for text-embedding-3-large, so the large model costs 6.5x as much per token. OpenAI's pricing page lists no separate Batch rate for these embedding models.
How is Gemini Embedding 2 priced for images, audio and video?
Gemini Embedding 2 bills each input type per 1M tokens: text $0.2, image $0.45, audio $6.5, video $12. The Batch API halves each rate.
Do embedding APIs charge for output tokens?
No. All models in this comparison bill only the input tokens you send. The returned vectors are not billed as output tokens, which is why the table shows a single input rate per model.
Which embedding APIs offer a Batch discount?
In this comparison, Gemini Embedding and Gemini Embedding 2 have published Batch rates. Batch suits indexing a corpus, where results can wait; search queries need the standard real-time rate.
Is embedding usually a big part of a RAG budget?
Usually not. Embedding a corpus is a one-time cost plus small monthly re-embedding and query costs, and per-token rates are far below LLM generation prices. In most retrieval-augmented generation (RAG) apps, the LLM calls that use the retrieved context and the vector database cost more than the embeddings.
Where do these prices come from?
Each price is a verified record from the AICostBudget pricing dataset, linked to the provider's official pricing page with its verification date. Multimodal inputs: Cohere Embed 5 Fast, Cohere Embed 5 Pro, Gemini Embedding 2.