BroadRouter

Embedding cost calculator

Updated

In short: Embedding cost = tokens embedded ÷ 1,000,000 × price per million tokens, paid once to index, again for changed documents and re-indexes, and for every search query. One million documents of 500 tokens is 500 million tokens: USD 10 on text-embedding-3-small, USD 65 on text-embedding-3-large and USD 75 on gemini-embedding-001 (list prices, 2026-11-14).

Free, no sign-up, runs in your browser: nothing you type is sent anywhere.

Corpus and traffic

Embedding cost by model
ModelPer 1M tokensInitial indexPer monthFirst year

BroadNet Inference lists no embedding model in its pricing.json today (synced 2026-11-14). If it adds one, it appears in this table automatically.

Get started

Estimates at list prices, before tax, checked 2026-11-14. They ignore free tiers (Voyage's 200 million free tokens, Google's free tier), volume discounts, and vector database storage and query fees, which are often the larger recurring cost. Words and pages are converted at 1,290 tokens per 1,000 English words; each embedding model has its own tokenizer, so measure a sample with the provider's usage field before you commit.

Report a wrong price or number (email to info@broadnet.ai, with this page's link and your inputs filled in).

How it works

  1. Describe the corpus: documents × average tokens, documents × average words, or pages. Words and pages are converted with the English ratio from our open tokenizer dataset.
  2. Set chunk overlap, the share of the corpus that changes each month and how many full re-indexes you expect a year (for example after switching models).
  3. Enter search queries a month and their average length in tokens. Tick "Use batch prices" if you index through a batch API.
  4. Read the table: initial indexing, monthly and first-year cost for each embedding model. Copy the link or download CSV.

The formula

corpus tokens   = documents × tokens per document
                  (words: documents × words × 1,290 / 1,000;
                   pages: pages × words per page × 1,290 / 1,000)
embedded tokens = corpus tokens / (1 − overlap)
initial index   = embedded tokens / 1,000,000 × price
per month       = embedded tokens × (changed share + re-indexes / 12) / 1,000,000 × price
                + queries × query tokens / 1,000,000 × query price
first year      = initial index + 12 × per month

Overlap matters because each chunk repeats the end of the one before it: with 20% overlap you embed every token 1.25 times. "Use batch prices" applies only to indexing, since search queries need an answer at once. The 1,290 figure is the mean of five real tokenizers on the English texts in our tokenizer efficiency dataset. See embeddings, token and batch API in the glossary.

Embedding API prices, compared

List prices in USD per 1 million input tokens, from each provider's own pricing page on 2026-11-14.

Provider and modelPer 1M tokensBatch, per 1MNotes and source
OpenAI text-embedding-3-smallUSD 0.02USD 0.01Up to 8,192 tokens per input. 1,536 dimensions by default. Batch API is 50% off. Official pricing, checked 2026-11-14.
OpenAI text-embedding-3-largeUSD 0.13USD 0.065Up to 8,192 tokens per input. 3,072 dimensions by default. Batch API is 50% off. Official pricing, checked 2026-11-14.
Google gemini-embedding-001USD 0.15USD 0.075Gemini API paid tier; Google lists it as free of charge on its free tier. Official pricing, checked 2026-11-14.
Google gemini-embedding-2 (text)USD 0.2USD 0.1Multimodal model; text price shown. Free of charge on the free tier. Official pricing, checked 2026-11-14.
Voyage AI voyage-4-largeUSD 0.12USD 0.08First 200 million tokens free per account (not applied here). Batch API is 33% off. Official pricing, checked 2026-11-14.
Voyage AI voyage-4USD 0.06USD 0.04First 200 million tokens free per account (not applied here). Batch API is 33% off. Official pricing, checked 2026-11-14.
Voyage AI voyage-4-liteUSD 0.02USD 0.01333First 200 million tokens free per account (not applied here). Batch API is 33% off. Official pricing, checked 2026-11-14.
Mistral mistral-embedUSD 0.1USD 0.05Cached tokens listed at USD 0.01 per 1M. Mistral's model page lists batching for this model and its batch docs give a 50% discount. Official pricing, checked 2026-11-14.
Mistral codestral-embed-2505USD 0.15USD 0.075Code embeddings. Mistral's model page lists batching for this model and its batch docs give a 50% discount. Official pricing, checked 2026-11-14.
Jina AI jina-embeddings v3 / v4 / v5USD 0.05No batch pricePrepaid token packs, not a per-token rate: 1 billion tokens for USD 50 (USD 0.045 per 1M on the 11-billion pack). Read from Jina's API dashboard product list; one balance covers all Jina models. Official pricing, checked 2026-11-14.
Cohere Embed v4 / Embed 5Not publishedNot publishedcohere.com/pricing listed no per-token Embed price on 2026-11-14; Cohere's docs point to that page. Ask Cohere sales for a quote. Pricing page.

text-embedding-3-small and voyage-4-lite are the cheapest per token here, and OpenAI's batch price halves its list price. The cheapest model is not always the cheapest system: a model with more dimensions stores bigger vectors, and a weaker model can force you to retrieve more chunks into every LLM prompt. BroadNet Inference does not sell embeddings today; it serves the chat models that answer over your index, priced from USD 0.1 per 1M input tokens.

Worked example: a support knowledge base

20,000 help articles of 800 words each is 16 million words, or about 20.6 million tokens at 1,290 tokens per 1,000 words. Indexing it once on text-embedding-3-small costs USD 0.41, and on text-embedding-3-large USD 2.68. 200,000 searches a month of 30 tokens add 6 million tokens: USD 0.12 or USD 0.78 a month. At this size embeddings are a rounding error next to the LLM calls that answer each question, which the LLM API cost calculator prices.

How to keep embedding costs down

  • Hash each chunk and re-embed only chunks whose text changed.
  • Keep overlap small; 10% to 15% is a common starting point, and every extra point is paid on the whole corpus.
  • Index through a batch API when you can wait for results.
  • Pick the model once, on a test with your own queries: switching later means a full re-index.
  • Strip boilerplate (navigation, footers, repeated disclaimers) before chunking. Count what is left with the token counter.

Questions

How much does it cost to embed 1 million documents?

Multiply the documents by their average length in tokens, then by the price per million tokens. One million documents of 500 tokens is 500 million tokens: USD 10.00 on text-embedding-3-small, USD 65.00 on text-embedding-3-large and USD 75.00 on gemini-embedding-001, at list prices checked 2026-11-14. Batch APIs cut that by a third to a half.

How much do OpenAI embeddings cost?

OpenAI lists text-embedding-3-small at USD 0.02 and text-embedding-3-large at USD 0.13 per million tokens (checked 2026-11-14). Its Batch API, which accepts the embeddings endpoint, is 50% off: USD 0.01 and USD 0.065. Both models take up to 8,192 tokens per input.

Which embedding API is cheapest?

On list price per million tokens, text-embedding-3-small and voyage-4-lite tie at USD 0.02; the lowest batch price in our table is USD 0.01. Price is rarely the deciding factor, though: at these rates, retrieval quality on your own data and the vector storage that grows with dimensions usually matter more. Test two or three models on 100 of your real queries.

Do you pay to embed search queries?

Yes. Every search query is embedded with the same model as the documents, and billed the same way. Queries are short, so the bill is usually small: 1 million queries of 30 tokens is 30 million tokens a month. Enter your query volume to see it.

When do you need to re-embed documents?

Whenever a document changes, re-embed its changed chunks. When you switch embedding models, or a provider retires one, re-embed everything: vectors from different models are not comparable, so the old index cannot be mixed with the new one. Set "full re-index per year" to budget for model upgrades.

How many tokens is a page of text?

At 500 words a page, a page of English prose is about 645 tokens: our open tokenizer dataset measures 1,290 tokens per 1,000 English words on average. Arabic, code and tables need more. Embedding models use their own tokenizers, so treat this as an estimate. Convert tokens and words.

Go deeper

Read as MarkdownAsk ChatGPTAsk ClaudeAsk Perplexity