CalcSays
EMBEDDING COST · CHURN, NOT JUST ONE-TIME

Embedding Cost Calculator

Embedding your corpus looks like a one-time cost. It isn't — churn means you're re-embedding forever, and switching models means re-embedding everything. This calculator shows the real first-year number.

For engineers budgeting a vector index — shows that the embed API charge is a decoy; vector storage, corpus churn, and chunk overlap are what actually recur, month after month.

Model prices from OpenRouter · updated 2026-07-13

01Your corpus & index

Embedding model
Vector database
Batch the initial embedOpenAI: 50% off

02 One-time vs real TCO

Real first-year cost is 12.1× the one-time number — storage is 82% of the recurring bill.

One-time (naive)
$0.59
embed the corpus once, batch rate
Real first-year
$7.14
+ storage, churn & queries, 12mo

$0.59 one-time + $5.37 storage + $0.71 churn + $0.48 queries = $7.14

147,059 vectors × 1536 dims → 1.355 GB on Pinecone, forever.

03 Cheapest combo, first-year

1. voyage-4-liteVoyage AI · cheapest$3.76 first-yr
2. text-embedding-3-smallOpenAI$7.14 first-yr
3. voyage-4Voyage AI$13.08 first-yr
4. embed-4Cohere$17.75 first-yr
5. voyage-4-largeVoyage AI$19.00 first-yr
6. text-embedding-3-largeOpenAI$22.27 first-yr
7. Gemini Embedding 001Google$24.04 first-yr
8. Gemini Embedding 2Google$28.48 first-yr
Tweet
Why isn't this a one-time cost?Storage is billed per vector, every month, and vectors don't shrink — it typically overtakes the one-time embed charge within a few months. Corpus churn and query embedding add smaller recurring amounts. A model or dimension migration re-embeds and re-stores everything, since indexes can't mix embedding versions.
📄 Get the full agent cost audit
A multi-page report on your exact setup: where cost spirals, how much context-trimming / state-resets / model-routing each saves, and which model+framework combo is cheapest at your real failure rate. Generated from your numbers.

Embedding prices hand-verified (checked 2026-07-13): OpenAI pricing · Google pricing · Voyage AI pricing · Cohere pricing. Vector storage hand-verified (checked 2026-07-13); index overhead estimated at 1.5× raw vector size. Both re-audited quarterly — neither is in the daily-synced catalog. Chunks = ceil(corpus ÷ (chunkTokens × (1−overlap))). All math runs in your browser.

How the math works

Embedding a corpus looks like a one-time API charge: 147,059 chunks of 400 tokens (with 15% overlap) cost $0.59 to embed on text-embedding-3-small (batch rate). For a typical corpus that's a few dollars — which is exactly why a plain 'embedding cost calculator' misses the point.

The real, recurring bill is storage, and it's billed per VECTOR, not per token: 147,059 vectors at 1536 dimensions need 1.355 GB on Pinecone (incl. ~1.5× index overhead for the live graph) — $0.45/month, forever. At this page's defaults storage is 82% of the recurring bill; churn and query embedding are the rest.

Same baseline, one identity: one-time $0.59 + storage (12mo) $5.37 + churn (12mo) $0.71 + queries (12mo) $0.48 = $7.14 real first-year cost — 11.1× the naive number. Storage costs typically catch up to and pass the one-time embed charge within a few months, then keep billing forever.

Chunk overlap (15% here) is a second, separate underestimate: it inflates the actual billed tokens (and vectors stored) by roughly the overlap fraction — 58.8M billed vs 50.0M in the raw corpus. And a model or dimension migration re-embeds and re-stores everything from scratch, since indexes can't mix embedding versions — budget that like a second one-time cost, not a delta.

Vector storage and embedding prices aren't in the daily-synced catalog; both are hand-verified against provider docs (embeddings checked 2026-07-13, vector DB storage checked 2026-07-13) and re-audited quarterly. All math runs in your browser.

Frequently asked questions

What does it really cost to run your corpus?

The one-time embed is $0.59 on text-embedding-3-small. But storing 147,059 vectors on Pinecone costs $0.45/month — forever — and at this page's defaults the real first-year cost is $7.14, 1114% above the one-time number. Tune the sliders for your real numbers.

Why does vector storage cost more than embedding?

Embedding is billed per token, once (or per churned token, monthly). Storage is billed per vector, every month, and vectors don't shrink. At this page's defaults storage is 82% of the recurring bill — for any corpus that lives more than a few months, the storage line eventually dwarfs everything else. A calculator that only shows the embed API charge is answering the wrong question.

How much does chunk overlap really cost?

At 15% overlap and 400-token chunks, your corpus is billed (and stored) as 58.8M tokens instead of the raw 50.0M — both the embed charge and the vector count inflate together, since more overlap means more, denser chunks. Higher overlap improves retrieval at chunk boundaries, but it's not free on either bill.

What does a model or dimension migration really cost?

The full one-time cost again — $0.59 at this corpus size — plus a fresh storage baseline, because embeddings from different models or dimensions aren't compatible in the same index. There's no partial migration; budget a switch like standing up the index from scratch.

Which model + vector DB combo is cheapest?

At this workload: voyage-4-lite leads at $3.76 first-year, then text-embedding-3-small ($7.14) and voyage-4 ($13.08). Retrieval quality (and your vector DB choice, which usually matters more than the embedding model) should decide this before price does.

What formula does this calculator use?

Chunks = ceil(corpusTokens ÷ (chunkTokens × (1 − overlap))); effective tokens = chunks × chunkTokens. One-time = effective tokens × embed price (batch rate if selected). Storage = chunks × dimensions × 4 bytes × 1.5 index overhead, in GB × $/GB. Monthly recurring = storage + churn × effective tokens × price + queries × query tokens × price. First-year = one-time + 12 × recurring.

Are these prices current?

Neither embedding nor vector-DB storage prices are in our daily-synced OpenRouter catalog — both are hand-verified directly against provider pricing pages (embeddings 2026-07-13, storage 2026-07-13) and re-audited quarterly. All math runs client-side with tested code.