Embeddings DB cost by stored vectors & query volume
Short answer: for a RAG index the monthly bill is driven by how many vectors you store and how many searches you run, and the cheapest vendor changes as you scale. Across this 42-cell sweep of 768-dim vectors, the cheapest managed option ranges from $1.57 to $39.36/mo, and only 2 of the four priced vendors (Turbopuffer and LanceDB) ever win. At 10M vectors and 250k queries/day the floor is $5.08/mo on Turbopuffer. The full grid is below.
Cheapest monthly cost by stored vectors and queries per day
Rows: stored vector count. Columns: queries per day. Each cell is the cheapest of four vendors' monthly cost, with the winning vendor noted. Green < $5, grey $5–$20, amber $20–$40, red ≥ $40.
| Vectors \ q/day | 10k | 50k | 100k | 250k | 500k | 1M | 2M |
|---|---|---|---|---|---|---|---|
| 1M | $1.57Turbopuffer | $1.62Turbopuffer | $1.68Turbopuffer | $1.86Turbopuffer | $2.16Turbopuffer | $2.76Turbopuffer | $3.96Turbopuffer |
| 5M | $3.00Turbopuffer | $3.05Turbopuffer | $3.11Turbopuffer | $3.29Turbopuffer | $3.59Turbopuffer | $4.19Turbopuffer | $5.39Turbopuffer |
| 10M | $4.59LanceDB | $4.84Turbopuffer | $4.90Turbopuffer | $5.08Turbopuffer | $5.38Turbopuffer | $5.98Turbopuffer | $7.18Turbopuffer |
| 25M | $5.39LanceDB | $10.20Turbopuffer | $10.26Turbopuffer | $10.44Turbopuffer | $10.74Turbopuffer | $11.34Turbopuffer | $12.54Turbopuffer |
| 50M | $6.73LanceDB | $12.13LanceDB | $18.88LanceDB | $19.38Turbopuffer | $19.68Turbopuffer | $20.28Turbopuffer | $21.48Turbopuffer |
| 100M | $9.41LanceDB | $14.81LanceDB | $21.56LanceDB | $37.26Turbopuffer | $37.56Turbopuffer | $38.16Turbopuffer | $39.36Turbopuffer |
Provenance
- Engine
- Embeddings DB Cost (
embeddings-db-cost) - Source
- Computed live from
/engines/embeddings-db-cost.js - Grid
- 6 vector counts × 7 query volumes = 42 cells
- Computed
- 2026-05-23
- Held constant
- 768-dim vectors · 20,000 ingests/day · 180-day retention · list pricing 2026-04 for Pinecone, pgvector, LanceDB, Turbopuffer
Every value above and in the table is the deterministic return value of the shipped engine bundle, recomputed independently in continuous integration and diffed against this page on every build. The engine is pure (no clock, no randomness): the same inputs always produce the same output. No number on this page was hand-typed or estimated.
How to read this dataset
- Find your stored vector count on the left, your daily query volume across the top, and read the cheapest monthly cost and winning vendor where they meet.
- The cheapest vendor flips as you scale: a self-hosted-style option wins on storage-heavy, query-light indexes, while a per-operation vendor wins when reads dominate.
- The CSV carries every vendor's full monthly price per cell plus the index storage in GB, so you can compare like-for-like and add your own compute overhead.
Want to run your own inputs instead of this grid? Use the Embeddings DB Cost calculator. For the exact math, defaults, and what the model can't tell you, read the methodology page. For a head-to-head at 200k vectors, see pgvector vs. Pinecone at 200k vectors, and for the build-vs-buy call, managed vector DB vs. self-hosted pgvector.