A vector DB bill comes from three numbers, not from the model. How many vectors you keep, how wide each one is (its dimension), and how many queries you run. Everything below is derived from those three — storage is vectors × dimension × 4 bytes plus index overhead, and query cost scales with read volume.
Monthly cost by provider
Same workload, ranked cheapest first. Storage + query estimate, before any free tier.
| Provider | Model | Est. / mo |
|---|
How the storage number is built
Each vector is an array of 32-bit floats, so a single 1536-dimension embedding is about 6 KB raw (1536 × 4 bytes). Multiply by your vector count for the raw payload, then add roughly 50% overhead for the HNSW/IVF index structure and metadata that every serious vector store keeps alongside the raw vectors. That's the "indexed storage" figure above — the thing every provider charges against, whether they call it storage, a pod, or a cluster size. Doubling the dimension (1536 → 3072) doubles this number for the same vector count, which is why "bigger embedding = better" is not a free choice.
Why the providers price so differently
Pinecone and Zilliz Cloud are usage-based serverless: you pay for stored GB plus read and write units, so a read-heavy RAG endpoint can cost more than its storage suggests. Qdrant Cloud and Weaviate Cloud are closer to compute/RAM-based — you pick a cluster size that must hold the index in memory, so cost tracks storage more than queries. pgvector is just an extension inside Postgres: if you already run a database, adding vectors is nearly free until the index outgrows RAM, at which point you pay for a bigger box. The calculator flattens all of these into one storage + query shape so you can compare order-of-magnitude, then you confirm the exact bill on the provider you pick.
How to keep the bill down
Store fewer, better chunks rather than every paragraph twice. Use a smaller embedding dimension when recall allows — 768-d often matches 1536-d on narrow domains at half the storage. Drop stale vectors instead of letting the index grow forever. And remember the read side: caching frequent queries and the LLM context around them is where the real money is. Pair this with the prompt caching savings calculator and, if you're sizing a whole retrieval pipeline, the chatbot cost calculator or the full AI app cost estimator.
Picking embeddings? The OpenAI and Gemini guides list current embedding prices, and the AI agent cost calculator covers retrieval-augmented agents end to end.