Every vendor meters differently, so the sticker prices don't line up. Here's roughly what 10 million vectors with steady query traffic costs on each managed cloud:
| ~10M vectors | Model | Self-host? | |
|---|---|---|---|
| Qdrant Cloud | ~$65/mo | per-node, no per-query fee | yes — Apache 2.0 core, free |
| Pinecone Serverless | ~$70/mo (+ $50/mo plan minimum) | storage $0.33/GB + read units + write units | no — fully managed only |
| Weaviate Cloud | ~$135/mo (or ~$27 with binary quantization) | storage + dimensions processed | yes — BSD core, free |
Qdrant is the cheapest managed option at scale and its open-source core is genuinely production-grade — a lot of teams self-host it on a $50–100/month box and skip the cloud entirely. Written in Rust, fast, good filtering.
Pinecone invented "vector database" as a product category and it shows in the polish: serverless scaling, zero ops, the smoothest getting-started experience. The costs are a $50/month plan floor plus metered storage, reads and writes — predictable once you understand the three dials, opaque until you do. There's no self-host option, so you're committed to the platform.
Weaviate is the most expensive per million vectors unless you turn on quantization — binary quantization cuts memory (and cost) around 5x with a small recall hit, and with it on, Weaviate becomes competitive. It also bundles more: hybrid search, built-in embedding and reranker modules, multi-tenancy.
For most RAG projects: start with Qdrant (self-hosted if you have the ops capacity, cloud if not) for the best cost. Choose Pinecone if you want zero operational burden and the plan minimum doesn't bother you. Choose Weaviate if you'll use its built-in modules and you enable quantization from the start.