How much does it cost to embed my documents?

Usually less than people expect. 10,000 pages, about 6.7 million tokens, cost from 13 cents on voyage-4-lite to $1.33 on Gemini Embedding 2 to embed once, among the newest models. An embedding model is priced only for the tokens it reads.

Does chunk overlap make embedding dearer?

Yes, in proportion. When half of each chunk repeats the one before, as with OpenAI's file search default of 400 tokens of overlap on chunks of 800, every token is embedded twice. The same 10,000 pages then cost 27 cents on voyage-4-lite instead of 13 cents.

How much storage do embeddings need?

Each vector is a list of numbers, usually stored at 4 bytes each. 10,000 pages in chunks of 800 tokens make 8,334 vectors: 102 MB at 3,072 numbers a vector and 34.1 MB at 1,024, before the index adds its own structures.

Do I have to embed everything again if I change the embedding model?

Yes. Vectors from 2 different models cannot be compared, so a new model means embedding every chunk again into a new index. Keep the parsed text of your documents, so that a change of model is a rerun rather than a rebuild from the sources.

Is embedding or the answering model the larger cost of RAG?

For most assistants, the answering model. 10,000 questions a month, each sent with 5 passages of 800 tokens and answered in a paragraph, cost $6.33 on Claude Haiku 5.5, while embedding those questions to search with costs 0.6 cents on voyage-4-lite. The model that answers reads thousands of tokens for each question, at a higher price per token, and writes the answer too.

How much does a RAG system cost a month?

Mostly what the answering model costs. With 10,000 pages, 5% of them changing each month and 10,000 questions, embedding costs 1.3 cents a month on voyage-4-lite, and answering costs from $6.33 on Claude Haiku 5.5 to $633 on GPT-6 Astra, among the newest chat models. Fewer or shorter passages cut the answering bill the most.

Which embedding model should I choose?

Not on price alone, since for most collections the cheapest and the dearest differ by a few dollars. Try 2 or 3 on your own real questions and check which finds the passage that answers each one. Then price the one that wins here.

What a RAG system costs, from embedding to answer

An assistant that answers from your documents pays twice: once to turn every document into vectors and keep them up to date, and again for each question, when a chat model reads the passages found and writes the answer. Say how much text you have and how many questions come in, and see both bills.

Price your documents Token counter Every price

Your documents

pages

6,666,667 tokens

New or changed each month

Embedded again, so the index keeps up with the documents.

Questions

a month

Each question is embedded to search with, at about 30 tokens, then answered by the chat model chosen under Answering.

The documents can waitThe batch price for the documents, where a maker has one. Questions need an answer at once.

Each prompt is 4,330 tokens: 5 chunks of 800, the question and 300 tokens of instructions. The answer is about 400 tokens.

The answer

Embedding 10,000 pages once costs from 13 cents on voyage-4-lite to $1.33 on Gemini Embedding 2. The vectors take 34.1 MB to 102 MB to store. Answering 10,000 questions a month on Claude Haiku 5.5 costs $6.33, 0.063 cents a question.

Embedding a month: 1.3 cents; voyage-4-lite
Answering a month: $6.33; Claude Haiku 5.5
Together a month: $6.34; After the first embedding

Maker

The first embedding, cheapest first

voyage-4-liteVoyage AI
13 cents
1.3 cents a month after that · $0.02 per million tokens · 8,334 vectors of 1,024 numbers, 34.1 MB
text-embedding-3-smallOpenAI
13 cents
1.3 cents a month after that · $0.02 per million tokens · 8,334 vectors of 1,536 numbers, 51.2 MB
voyage-4Voyage AI
40 cents
3.8 cents a month after that · $0.06 per million tokens · 8,334 vectors of 1,024 numbers, 34.1 MB
voyage-4-largeVoyage AI
80 cents
7.6 cents a month after that · $0.12 per million tokens · 8,334 vectors of 1,024 numbers, 34.1 MB
voyage-context-4Voyage AI
80 cents
7.6 cents a month after that · $0.12 per million tokens · 8,334 vectors of 1,024 numbers, 34.1 MB
Embeds each chunk with the context of its whole document.
voyage-code-4Voyage AI
80 cents
7.6 cents a month after that · $0.12 per million tokens · 8,334 vectors of 1,024 numbers, 34.1 MB
Made for code.
text-embedding-3-largeOpenAI
87 cents
8.2 cents a month after that · $0.13 per million tokens · 8,334 vectors of 3,072 numbers, 102 MB
Gemini Embedding 2Google
$1.33
13 cents a month after that · $0.20 per million tokens · 8,334 vectors of 3,072 numbers, 102 MB

Voyage AI: The first 200 million tokens on an account are free on the newest models, but not through batch.Google: The standard API has a free tier, with lower limits.List prices, no discounts or free allowances taken off. Storage counts 4 bytes a number, before any index or compression; what it costs depends on where you keep it.

The memory and disk these 8,334 vectors need with an index

Answering, cheapest first

Each of the newest chat models reading 5 passages and writing the answer. Choose one to count it in the month above.

List prices for answers given at once, with no cache and no thinking tokens. A model that thinks first writes more, and a cached instruction block costs less. Price one model in full.

The makers' own pages

How the sum works

An embedding model reads text and returns a list of numbers. It writes no text, so it is priced only for what it reads, per million tokens. The tool cuts your documents into chunks of the size you choose and adds up the tokens each model reads.

Overlap repeats the end of each chunk at the start of the next, so a sentence cut at a boundary still appears whole in one of them. Those tokens are embedded twice: a quarter of overlap adds a third to the bill, and half doubles it.

After the first run, documents change. The share you set is embedded again each month, and so is each question people ask, at about 30 tokens, to search with. With "The documents can wait", the documents use the batch price where a maker has one: Google prints a batch price, and Voyage AI takes a third off. Questions always pay the standard price, because they need an answer at once.

Storage counts the vectors alone, at 4 bytes a number. A vector index adds its own structures, and each database charges in its own way, so treat it as the least you will store. Many models can return shorter vectors, which take less space; whether the shorter ones find the right passages as often is worth testing on your own questions.

Answering is the second bill. For each question, the passages the search found go to a chat model with the question and the instructions, and the model writes the answer. The tool counts the prompt as the passages you choose, each one chunk long, plus the question and 300 tokens of instructions, and prices it at each model's standard input and output prices, with the long-prompt price where a maker charges more past a length. It leaves out caching, batch prices and thinking tokens; the cost calculator prices those for one model.

Free allowances are not taken off, only shown with each model. Anthropic does not make an embedding model, and its documentation points to Voyage AI. If a figure here and the maker's page disagree, the maker's page is right.

Questions

How much does it cost to embed my documents?: Usually less than people expect. 10,000 pages, about 6.7 million tokens, cost from 13 cents on voyage-4-lite to $1.33 on Gemini Embedding 2 to embed once, among the newest models. An embedding model is priced only for the tokens it reads.
Does chunk overlap make embedding dearer?: Yes, in proportion. When half of each chunk repeats the one before, as with OpenAI's file search default of 400 tokens of overlap on chunks of 800, every token is embedded twice. The same 10,000 pages then cost 27 cents on voyage-4-lite instead of 13 cents.
How much storage do embeddings need?: Each vector is a list of numbers, usually stored at 4 bytes each. 10,000 pages in chunks of 800 tokens make 8,334 vectors: 102 MB at 3,072 numbers a vector and 34.1 MB at 1,024, before the index adds its own structures.
Do I have to embed everything again if I change the embedding model?: Yes. Vectors from 2 different models cannot be compared, so a new model means embedding every chunk again into a new index. Keep the parsed text of your documents, so that a change of model is a rerun rather than a rebuild from the sources.
Is embedding or the answering model the larger cost of RAG?: For most assistants, the answering model. 10,000 questions a month, each sent with 5 passages of 800 tokens and answered in a paragraph, cost $6.33 on Claude Haiku 5.5, while embedding those questions to search with costs 0.6 cents on voyage-4-lite. The model that answers reads thousands of tokens for each question, at a higher price per token, and writes the answer too.
How much does a RAG system cost a month?: Mostly what the answering model costs. With 10,000 pages, 5% of them changing each month and 10,000 questions, embedding costs 1.3 cents a month on voyage-4-lite, and answering costs from $6.33 on Claude Haiku 5.5 to $633 on GPT-6 Astra, among the newest chat models. Fewer or shorter passages cut the answering bill the most.
Which embedding model should I choose?: Not on price alone, since for most collections the cheapest and the dearest differ by a few dollars. Try 2 or 3 on your own real questions and check which finds the passage that answers each one. Then price the one that wins here.

Put this calculator on your own site

Copy the code into any web page: a blog post, a course page or your team's wiki. The calculator runs in a frame and shows the same figures as this page, so when the prices change here they change on your page too. It is free and needs no account, and nothing a reader types is sent anywhere.

The code opens the calculator with the settings you have chosen above.

The short script at the end lets the frame grow and shrink with the calculator. Where a page cannot take scripts, choose a fixed height and the frame scrolls.

<iframe src="https://tecyfy.com/embed/embedding-cost" title="RAG and embedding cost" width="100%" height="3900" style="border:0;display:block" loading="lazy" allow="clipboard-write"></iframe>
<p><a href="https://tecyfy.com/prices/embedding-cost">RAG and embedding cost</a> by Tecyfy</p>
<script>addEventListener("message",function(e){if(e.origin!=="https://tecyfy.com"||!e.data||e.data.type!=="tecyfy:height")return;var h=Number(e.data.height);if(!(h>0))return;document.querySelectorAll("iframe").forEach(function(f){if(f.contentWindow===e.source)f.style.height=Math.min(h,20000)+"px"})})</script>

The embedding cost tool

The makers' own pages

How the sum works

Questions

Put this calculator on your own site