Vector database cost comparison, with the hidden fees.
Headline prices for Pinecone, Weaviate, Qdrant, and open-source vector stores hide most of what you'll actually pay. Here's how the components break down, what they cost at scale, and where CogniWeave fits.
What you actually pay for
- • Storage: bytes of vector + metadata kept on disk and in RAM.
- • Query volume: number of nearest-neighbour searches per month.
- • Embedding operations: tokens encoded at ingest and at query time.
- • Support & SLA tier: response time guarantees and uptime credits.
Cost comparison
- • Pinecone: serverless pay-per-query (~$0.04 per 1k queries + ~$1/month per index). No upfront commit.
- • Qdrant: self-hosted is free; Qdrant Cloud starts at ~$9/month with per-vector overage.
- • Weaviate: Cloud starter ~$25/month; self-hosted is free if you operate it.
- • Open-source vector extensions (self-hosted): ~$5–$20/month for small-to-medium workloads.
- • CogniWeave: ingestion, embedding, storage, and retrieval included in plan—see pricing page.
Worked example (10M vectors)
- • Pinecone serverless: $500–$2,000/month at typical RAG query volumes.
- • Weaviate Cloud: $200–$800/month depending on replica + tier.
- • Self-hosted open-source vector index: $50–$200/month plus engineering time.
- • CogniWeave: bundled within your subscription—no per-query overage at this scale.
When to choose what
- • Pinecone if you want zero ops and accept usage-based billing.
- • Weaviate if you want control and a mature open-source ecosystem.
- • An open-source vector extension if your corpus is small and you already run a relational database.
- • CogniWeave if you want ingestion, retrieval, and citations without integration work.
Common questions
What hidden costs should I watch for with managed vector databases?
Egress fees, per-namespace charges, replica multipliers, support tier upgrades, and per-query overage above included tiers. Always model 90-day costs at projected query volume, not the headline starter price.
How do costs scale with corpus size?
For HNSW indexes, RAM scales roughly linearly with vector count. A 10M vector corpus at 1536 dimensions needs ~60 GB of RAM if kept hot. Pinecone serverless reduces this by paging to object storage at the cost of higher cold-start latency.
Are egress fees a real concern?
Yes—if your application server lives in a different region from your vector store, every query response incurs cross-region transfer. CogniWeave colocates compute, storage, and embedding generation in the same region to avoid this.
What does CogniWeave's pricing include?
Ingestion (parse, OCR, chunk), embedding generation, vector storage, retrieval API, and citations. There are no per-query surcharges below your plan limit and no egress fees within the platform.