Short answer. Choose Pinecone for a fully managed service with predictable units. Choose Qdrant for an open-source engine you can self-host, with strong filtered search and a managed option. Choose Weaviate for hybrid search, built-in modules and mature multi-tenancy. At 10M vectors, Pinecone is easiest to estimate; Qdrant needs a resource quote; Weaviate can be cheapest at low query volume.
Side by side
| Dimension | Pinecone | Qdrant | Weaviate |
|---|---|---|---|
| Deployment | Managed, serverless, BYOC | Open source, cloud, hybrid, private | Open source, cloud, enterprise |
| Free tier | 2 GB, 2M writes, 1M reads | 0.5 vCPU, 1 GB RAM, 4 GB disk | 100K objects, 1 GB memory, 10 GB disk |
| Entry paid | $20/month Builder; $50 minimum Standard | Usage-based Standard | $45/month Flex |
| Storage rate | $0.33/GB-month | Resource-based | From $0.12/GiB-month |
| Hybrid search | Dense, sparse and full-text indexes | Dense, sparse and BM25 hybrid | BM25 plus vector with fusion |
| Filtering | Metadata filters and namespaces | Payload indexes and filters | Where filters and tenants |
| Self-host | No | Yes | Yes |
| Best for | Teams that want no ops | Portable, filter-heavy retrieval | Hybrid and module-heavy stacks |
Sources: Pinecone, Qdrant and Weaviate pricing pages and documentation, checked 20 November 2026. Rates vary by cloud, region and plan; treat the numbers as list prices, not quotes.
Managed vs self-hosted
Pinecone removes the infrastructure decision. You create an index, choose a cloud and region, and query it. Capacity is measured in read units, write units and storage. The trade-off is that you cannot run the engine yourself: the bring-your-own-cloud option puts Pinecone in your cloud account, but it is still Pinecone's software and control plane.
Qdrant's open-source edition is the default for teams that want portability. You run it on Kubernetes or VMs, size memory for the HNSW index and payload, and own backups and upgrades. Qdrant Cloud adds a managed path with a free tier, usage-based Standard and a Premium tier with SSO and private links; Hybrid Cloud runs in your infrastructure.
Weaviate is also open source, with a managed cloud that has a free sandbox, a Flex pay-as-you-go tier and a Premium tier. Its enterprise edition adds dedicated deployment and compliance controls. Self-hosting is a real option, but the operational surface is larger if you use its modules and multi-tenancy.
Hybrid search and filtering
Hybrid search matters when exact terms, product codes or names must survive alongside semantic similarity. Pinecone supports dense vectors, sparse vectors and full-text indexes in one query, with metadata filters and namespaces for partitioning. That is a strong fit for search over catalogues and support content where both meaning and exact identifiers matter.
Qdrant's 1.20 release added BM25 hybrid search without requiring sparse vectors, while still supporting sparse vectors for learned sparse retrieval. Payload indexes make filtered search fast, which is important for multi-tenant applications where every query is scoped to an account. Qdrant's query API lets you fuse prefetches and rerank in one request.
Weaviate combines BM25 and vector search with fusion, plus multi-tenancy and modules for embeddings and generation. If you want the database to own more retrieval, Weaviate reduces moving parts; if you want logic in your application, Qdrant or Pinecone is thinner.
Cost at 10M vectors
The following model assumes 10 million vectors at 1,024 dimensions, float32, about 61 GB of index storage after HNSW overhead, 10 million queries per month and 1 million reindex writes per month. It excludes egress, backups, support and taxes. It is a sanity check, not a quote.
| Item | Pinecone Standard | Qdrant Cloud | Weaviate Flex |
|---|---|---|---|
| Storage | 61 GB × $0.33 = $20 | Resource-based | 61 GiB × $0.12 = $7 |
| Reads / queries | 10M × $16/M = $160 | Included in cluster | Included in cluster |
| Writes / reindex | 1M × $4/M = $4 | Included in cluster | Included in cluster |
| Vector dimensions | Not separately billed | Not separately billed | 10.24B × $0.00465/M = $48 |
| Modeled total | ≈ $184/month | Calculator or quote | ≈ $55/month |
Pinecone's read-unit definition matters: confirm how many results a unit covers. Weaviate's dimension rate varies by index, compression and region; $0.00465 per million is the Flex starting rate, and the $45 minimum applies. Qdrant's cluster price depends on vCPU, RAM, disk and replication; 10M vectors at 1,024 dimensions need roughly 40 GB before overhead, so three nodes with 16–32 GB RAM each is a realistic start.
Latency
Do not choose a vector database from a vendor's latency chart. Query latency depends on the number of vectors, dimensions, top-k, filters, payload size, index warm-up, replication and whether the index is quantized. A filtered query on a cold shard can be an order of magnitude slower than an unfiltered query on a warm one.
Measure p50 and p95 with your own filters and payloads. Test a selective filter, a top-k of 50 with reranking and a multi-tenant query that must not cross accounts. Dedicated read nodes, hybrid cloud and replication exist for teams that need predictable latency.
Migration
All three store a vector plus an ID and metadata, so the move is mechanical. The risk is index configuration and query semantics. Export IDs, vectors, metadata and permissions; recreate the collection with the target distance metric, quantization and HNSW settings; replay a shadow query set and compare recall@k.
# Generic migration shape: read, transform, upsert, verify.
for batch in source.scroll(batch_size=1000):
points = [{
"id": row["id"],
"vector": row["embedding"],
"payload": {
"text": row["text"],
"tenant_id": row["tenant_id"],
"source": row["source"],
"updated_at": row["updated_at"],
},
} for row in batch]
target.upsert(collection="docs", points=points)
# Verify: same query, same filter, compare top-10 IDs and scores.
assert recall_at_k(source, target, queries, k=10) > 0.98
Map namespaces to collections or tenant keys, not separate clusters. Keep the old index readable for one release, dual-write new content, and cut reads over after recall and p95 pass. Re-tune hybrid fusion weights; they do not transfer.
Which one to choose
- Pinecone for a small platform team that values a managed SLA, predictable units and no index operations.
- Qdrant for teams that want open-source portability, strong payload filtering and a managed option later.
- Weaviate for hybrid search, multi-tenancy and a stack that wants embedding and generative modules close to the data.
- None of them if your corpus fits in Postgres with pgvector and your query volume is modest; fewer systems is an advantage.
FAQ
Which is cheapest at 10 million vectors?
It depends on query and write volume. At 10 million 1,024-dimension vectors and 10 million queries a month, a modeled Pinecone Standard bill is $184 and Weaviate Flex about $55 before backups. Qdrant Cloud is usage-based and must be quoted; self-hosting Qdrant trades the bill for infrastructure and operations.
Can I self-host Pinecone?
No. Pinecone is a managed service, with a bring-your-own-cloud option for enterprises that runs in your cloud account. Qdrant and Weaviate both offer open-source editions you can self-host.
Which has the best hybrid search?
All three support hybrid retrieval. Pinecone combines dense and sparse indexes; Qdrant added BM25 hybrid search and also supports sparse vectors; Weaviate fuses BM25 and vector search.
How hard is migration?
The data move is straightforward: export IDs, vectors, metadata and permissions, then re-create the index. The hard parts are re-tuning HNSW and quantization, remapping namespaces or tenants, and proving recall did not drop. Run a shadow query set before cutover.
Does vector database latency decide the choice?
Rarely on its own. Most RAG latency is model time, not vector search. Filtering, index warm-up, replication and payload size affect p95 more than the engine's published benchmark. Measure recall and p95 on your own data and filters.
Reviewed 20 November 2026. Rates and features were checked against Pinecone pricing, Qdrant pricing and Weaviate pricing. The 10M-vector model is an editorial estimate, not a vendor quote; confirm current rates and unit definitions before committing.