Skip to content

Embedding Cost Calculator

Turn a corpus size and a chunking strategy into the one-time index cost, the monthly cost of new documents and queries, and the cost of the re-index you will eventually do.

First-year embedding spend
$178.27

Initial index $21.60, then $7.73 a month plus re-indexing.

Chunks per document
2
Tokens embedded per document
864
Overlap tax (tokens paid for twice)
8.0%
Initial index — tokens
216,000,000
Initial index — cost
$21.60
New documents
$1.73 / month
Query embeddings
$6.00 / month
Re-index (2× / year)
$63.94 / year
First-year total
$178.27
What this assumes: one embedding call per chunk, priced per 1M input tokens; overlapping tokens are billed once per chunk they appear in, which is why the tokens embedded exceed the corpus; a re-index is priced against the corpus at its mid-year size (start + six months of growth), so a year of heavy growth costs more than this shows; queries are embedded fresh every time, with no cache; nothing here covers storing the vectors — that is the vector storage calculator — and nothing covers the generation half of a RAG query.

Embedding a corpus is the one AI cost that is genuinely predictable, because there is no sampling in it: the same documents produce the same token count every time. What surprises people is not the initial number but how often they pay it again. Every change to the chunking strategy, every switch to a different embedding model, every migration to a different dimensionality re-embeds everything from scratch, and the re-index line above is usually larger than the initial index within eighteen months.

The overlap tax is the term most estimates drop. If you chunk at 512 tokens with 64 of overlap, roughly one token in eight is embedded twice — a 14% surcharge on the whole corpus, paid again on every re-index. That is not an argument against overlap, which exists because retrieval quality falls off a cliff when a sentence is split across a boundary. It is an argument for knowing what it costs before someone raises it to 128 to fix a retrieval complaint.

What this leaves out: the compute to parse, clean and de-duplicate the documents before they reach the embedding model, which for PDFs is frequently the larger bill; rate limits, which turn a $40 index into a three-day job; and failed batches that get retried in full. It also assumes one vector per chunk. If you store several representations of each chunk — a summary vector, a keyword vector, a per-section vector — multiply the document count by however many you keep.

Embedding Cost Calculator · Multigrid