What a vector database really costs: Pinecone vs Qdrant vs Zilliz vs pgvector

Published 2026-06-25 · reference numbers, verify before budgeting

Ask "how much does a vector database cost?" and you'll get four answers depending on who's selling. The honest answer is that for most retrieval-augmented (RAG) apps the bill is tiny until you scale — and that the storage layer is almost never the expensive part of an AI feature. The expensive part is the LLM call that happens after retrieval. But the vector DB is the line item people fixate on, so let's actually size it. We'll take one concrete workload and run it across Pinecone, Zilliz Cloud, Qdrant, Weaviate and self-hosted pgvector.

The workload we're pricing

A mid-size RAG app: 500,000 vectors, each a 1536-dimension embedding (the size OpenAI's text-embedding-3-small produces), serving 1,000,000 queries a month. That's a real product — a documentation search, a support copilot, an internal knowledge base with steady traffic. The first thing to compute isn't a price, it's the storage footprint, because that's what every provider charges against.

Step 1: the raw storage number

A vector is an array of 32-bit floats. So one 1536-dimension embedding is 1536 × 4 = 6,144 bytes, about 6 KB. Multiply by 500,000 vectors and that's ~3 GB of raw vectors. Add roughly 50% overhead for the index structure (HNSW or IVF graphs) and metadata every store keeps alongside the vectors, and you land near 4.6 GB of indexed storage. Hold that number — it's the whole story:

Prices are reference estimates, July 2026. Report outdated price →

Two things fall out of this immediately. First, dimension is a cost lever: switch to a 3072-dimension model (text-embedding-3-large) and you double this number for the exact same vector count. Second, 4.6 GB is small. That's the punchline most pricing pages bury.

Step 2: the monthly bill across providers

Here's where pricing models diverge, and why head-to-head tables are misleading. Pinecone and Zilliz Cloud bill usage — stored GB plus read/write units — so a query-heavy endpoint costs more than its storage implies. Qdrant and Weaviate bill compute/RAM — you pick a cluster big enough to hold the index in memory, so cost tracks storage, not queries. And pgvector is just an extension inside Postgres: if you already run a database, the marginal cost of adding vectors is close to zero. Flattened to a comparable monthly estimate for our 4.6 GB / 1M-query workload:

Look at the range: single-digit to low-double-digit dollars a month. Whichever way you cut it, the storage layer for a 500k-vector app is a rounding error next to the embedding and generation costs around it. The "which vector DB is cheapest" debate is real money only at tens of millions of vectors or extreme query rates — below that, pick on operations and features, not price.

Where the bill actually grows

Three things move a vector DB bill, and none of them is the logo on the dashboard:

• Vector count. Storage is linear in how many vectors you keep. Teams that re-embed the entire corpus on every document change, or never delete stale chunks, watch the index balloon. Embed only what changed.
• Dimension. 768-d often matches 1536-d on a narrow domain at half the storage and faster search. Bigger embeddings are not a free upgrade.
• Query volume — on usage-based providers. On Pinecone or Zilliz, 1M queries vs 10M queries is a real difference. On compute-based Qdrant or Weaviate, queries are "free" until you need a bigger box. This is the single biggest reason the cheapest provider depends on your read/write ratio.

So which should you pick?

• Already running Postgres, small-to-medium scale: pgvector. Near-zero marginal cost, one less service to operate, and it's plenty fast to a few million vectors. The trade-off is you own the indexing and scaling.
• Want managed, query-heavy, pay only for what you use: Pinecone or Zilliz serverless. No cluster to size; cost scales with traffic.
• Want predictable cost and full control of the index: Qdrant or Weaviate — you pick the cluster, the bill is flat, and queries don't meter.

The takeaway

For a 500k-vector, 1M-query RAG app in 2026, budget under ~$40/month for the vector database on any of these — and frequently a few dollars or effectively free on pgvector. Size the storage footprint first (vectors × dimension × 4 bytes × ~1.5), decide whether your traffic is storage-bound or query-bound, and then choose on operations rather than sticker price. The retrieval store is the cheap part of RAG; the LLM call after it is where you should spend your optimization energy — see the RAG cost calculator for the full picture.

Try the numbers yourself: the vector database cost calculator takes your vector count, dimension and query volume and compares all five providers. Then size the rest of the pipeline with the RAG cost calculator or the whole product with the AI app cost estimator. It's the same lesson as hidden API costs that double your bill: model the driver, not the logo.

Figures are reference estimates (June 2026) and change frequently. Vector DB pricing models differ substantially — usage-based vs compute-based — and free tiers (Pinecone Starter, Zilliz free cluster, Qdrant 1 GB free) are not subtracted here. Always confirm on each provider's official pricing page before budgeting.