Pricing report · 2026-08-22
We priced 21 vector databases so you don't have to
Same workloads, same rules, every vendor. The spread between the cheapest and the "default" choice at production scale is 9×, and the most expensive option isn't even the best-known one.
What moved this month
Two vendors made August an unusually good month to re-price the market. Astra DB left the ranking entirely. DataStax's IBM acquisition pulled its public serverless rates, so instead of quoting numbers we can no longer verify, we withdrew it until a live-verifiable offer exists. And Cloudflare Vectorize's billing model changed: it now bills per queried and stored vector dimension, not per query: a rate structure our engine models as included allowances only, because a flat per-query average would understate real bills by an order of magnitude. Both moves are documented in the catalog, not footnoted away.
Why another comparison
Most vector-database comparisons are feature matrices copied from marketing pages. Almost none of them tell you what a workload will actually cost, because pricing pages are full of capacity units (search units, OCUs, VCUs, nodes), regional multipliers, and free tiers that quietly expire.
We maintain a live comparison tool backed by a catalog of 21 vendors whose rates are hand-captured from official pricing pages and public APIs: each row carries a provenance date (20 of them stamped). This report runs all of them through three fixed cost-model workloads and reports USD-normalized monthly estimates, EU region. These are cost-model scenarios, not measured performance benchmarks: no latency, throughput, recall or availability equivalence was tested. Withdrawn from ranking: DataStax Astra DB Vector(DataStax's IBM acquisition pulled its serverless rates). Public pricing became unverifiable, so we refuse to quote numbers we cannot check. A vendor returns the moment a live-verifiable offer is captured.
The headline finding
At the mid-scale workload (1M vectors, 1M queries/month), prices range from $5.00/mo to $350.43/mo across the 19 quotable vendors. That's not a rounding difference. It's two different billing philosophies colliding:
- Metered serverless (Pinecone, Turbopuffer, Cloudflare Vectorize) bills per unit consumed, so small workloads stay near zero.
- Capacity floors (Azure AI Search search-units, AWS OpenSearch Serverless OCUs, Vertex AI nodes, Elasticsearch VCUs) bill for provisioned compute whether you query it or not.
The capacity-billed vendors aren't bad value, they're engineered for steady enterprise traffic where an SLA-backed floor is fine. But pointing a 100k-vector side project at any of them means paying $65.75–$350.42/month for what costs $0 elsewhere. At mid-scale, the floors look like this: Elasticsearch Serverless $65.78, Vertex AI Vector Search $68.48, Azure AI Search $73.73, AWS OpenSearch Serverless $350.43.
The numbers
RAG prototype
100,000 vectors · 1536-dim fp32 · 0.2M queries/mo
Mid-scale production
1,000,000 vectors · 768-dim int8 · 1M queries/mo
Large RAG
5,000,000 vectors · 768-dim int8 · 10M queries/mo
Walkthrough: size a one-million-vector RAG workload
Start with stored chunks, not uploaded document count: one chunk produces one vector in this example. Enter 1,000,000 vectors, 768 dimensions, INT8 precision, 500 metadata bytes per vector, index overhead enabled, 1 million queries and 0.1 million upserts per month, with the EU filter. These are the same inputs as the mid-scale table above; they are not inferred from traffic logs.
- Vector payload: count × dimensions × bytes per dimension. INT8 uses one byte; FP32 uses four. Divide bytes by 1,000,000,000 for decimal GB, not GiB.
- Index allowance: this model adds 50% of vector payload for the index. It is a planning assumption, not a measured HNSW memory requirement.
- Metadata: count × metadata bytes. Add this separately, rather than multiplying it by the index allowance.
- Price: apply the resulting footprint and monthly reads/writes to eligible plans. A capacity floor or allowance boundary can dominate the bill; halving bytes need not halve cost.
| Scenario | Vectors | Index | Metadata | Total |
|---|---|---|---|---|
| Baseline: INT8, 500-byte metadata | 0.768 | 0.384 | 0.500 | 1.652 |
| FP32 instead of INT8 | 3.072 | 1.536 | 0.500 | 5.108 |
| 2,000-byte metadata instead of 500 | 0.768 | 0.384 | 2.000 | 3.152 |
| Twice the queries, same stored vectors | 0.768 | 0.384 | 0.500 | 1.652 |
Baseline arithmetic: 0.768 GB vectors + 0.384 GB index allowance + 0.500 GB metadata = 1.652 GB. Moving to FP32 gives 5.108 GB, not four times the total, because metadata stays fixed. Doubling queries leaves storage unchanged but may increase metered charges or require more capacity. Monthly volume alone does not prove a node can handle burst traffic.
Use the live calculator to enter these inputs and inspect plan-level breakdowns. The link opens the calculator, not a prefilled scenario. Check provider billing units before budgeting: a request is not necessarily one read unit, and a vector-dimension meter is not a flat per-query rate. Quantization must also be validated for your recall requirements and supported deployment configuration.
Budget beyond the table: embedding generation, LLM inference, reranking, backups, replicas, network egress, taxes, migration and operator time are not a complete part of these estimates. A self-hosted VPS estimate is not a managed-service SLA. Test your own recall, latency and peak throughput before choosing a deployment; the cheapest modeled row is not a performance winner.
Three takeaways
- Free tiers are real now. Qdrant Cloud Managed and Zilliz Cloud (Managed Milvus) land at $0 on the prototype workload. Check eligibility, capacity and operational requirements before relying on a free tier; this does not make the whole RAG application free.
- The "default" choice carries a premium that grows with scale. Pinecone Serverless runs 1.8× the cheapest option at mid-scale (9× at the large workload, $85.43 vs $9.15). Sometimes the managed convenience is worth it. But it should be a decision, not a default.
- Self-hosting still anchors the market. pgvector on a $9.15 Hetzner VPS holds a top-3 spot at both mid-scale ($9.15/mo) and large workloads ($9.15/mo): the reason every managed vendor has to justify its margin.
Method & honesty notes
- Rates captured from official pricing pages / public APIs; capture date recorded per vendor row (
pricingCapturedAt). Azure figures cross-checked against Microsoft's Retail Prices API; Vertex node rates against Google's published us-central1 table. - Vendors that bill capacity rather than queries carry engineering-estimate query allowances, documented in their catalog rows, not hidden.
- EUR-denominated vendors are converted at the tool's fixed FX rate before ranking.
- Region-filtered ranking: US-only vendors (e.g. Chroma Cloud) are excluded from these EU-workload tables, which is why the count reads fewer than 21 where applicable.
- Originally published 2026-08-22; calculations on this page use the catalog bundled with the current build, whose full-verification date is 2026-08-22. This is not an immutable archive of August prices. A code or prose update is not a new price verification. Consult the capture dates and provider sources below; neither this page nor the live tool guarantees today's quote.
Source trail and freshness
These links are the catalog's recorded pricing sources, not proof that every rate was rechecked today. Capture dates can be older than the catalog-wide date. A withdrawn row remains excluded even if its source URL loads.
- Pinecone Serverless — captured 2026-08-20.
- Pinecone Pod-Based (Dedicated) — captured 2026-08-20.
- Qdrant Cloud Managed — captured 2026-08-20.
- Weaviate Cloud (WCS) — captured 2026-08-20.
- Zilliz Cloud (Managed Milvus) — captured 2026-08-20.
- Chroma Cloud Managed — captured 2026-08-20.
- DataStax Astra DB Vector — captured date unavailable; pricing needs review, excluded.
- Cloudflare Vectorize — captured 2026-08-20.
- Supabase Vector (pgvector) — captured 2026-08-20.
- Neon Serverless Postgres (pgvector) — captured 2026-08-20.
- pgvector on Hetzner VPS (Self-Hosted) — captured 2026-08-20.
- AWS RDS PostgreSQL (pgvector) — captured 2026-08-20.
- Turbopuffer — captured 2026-08-20.
- MongoDB Atlas Vector Search — captured 2026-08-20.
- Redis Cloud (Vector Search) — captured 2026-08-20.
- ClickHouse Cloud — captured 2026-08-20.
- Vespa Cloud — captured 2026-08-20.
- Azure AI Search — captured 2026-09-21.
- Elasticsearch (Elastic Cloud) — captured 2026-09-21.
- Google Vertex AI Vector Search — captured 2026-09-21.
- AWS OpenSearch Serverless — captured 2026-08-22.