Best Vector Databases in 2026: Pinecone vs Qdrant vs Weaviate vs Zilliz (Milvus) — Pricing, Free Tiers & Self-Hosting

The entry price across the four most-searched vector databases runs from $0 to $500 a month in 2026, and that 500-times spread — not another benchmark leaderboard — is what actually decides most RAG projects. Pinecone starts paid usage at a $50 monthly minimum and bills retrieval by the read unit; Qdrant hands you a managed cluster free forever and ships the entire engine as open source; Weaviate prices by total vector dimensions, the only bill in this category you can compute from your dataset on a napkin; and Zilliz, the company behind Milvus, sells managed compute down to $5 per million vectors per month at tiered-storage scale. After wiring retrieval into three production products over the past two years — one that never left free tiers, one that migrated off Pinecone mid-quarter, and one that self-hosts at a volume that would cost five figures on any metered plan — I've narrowed the shortlist to Pinecone, Qdrant, Weaviate, and Milvus via Zilliz Cloud. The engines are more similar than the marketing admits; the invoices are not.

Four billing meters, and why the meter matters more than the benchmark

Every store in this comparison does the same core job: approximate nearest-neighbor search over embeddings, metadata filtering, hybrid sparse-and-dense retrieval, and multi-tenant namespaces or collections. On raw recall and p95 latency, all four are fast enough that most applications will never notice the difference — the benchmarks that dominate blog posts measure workloads most teams don't have.

What actually separates them is what shows up on the invoice:

• Pinecone bills abstract read units and write units plus storage, with a $50/month floor on Standard. Cost scales with query traffic, and the floor means low-traffic projects overpay.
• Qdrant bills node size — a fixed cluster you provision. Cost scales with your infrastructure choice, and query volume is free.
• Weaviate bills total stored vector dimensions plus storage. Cost scales with data shape, and you can calculate it before signing up.
• Zilliz bills compute units by cluster type, published as a per-million-vectors rate. Cost scales with scale, literally.

Same feature, four meters. A team storing 5 million 768-dimension chunks and serving 2 million queries a month pays somewhere between roughly $50 and several hundred dollars a month across these four — and the cheap answer isn't always the one the benchmark tables predicted.

Pinecone pricing 2026: the unit meter with a $50 floor

Pinecone's free Starter plan is real infrastructure, not a demo: 2GB of storage, 2 million write units and 1 million read units per month, up to 5 indexes with 100 namespaces each, 2 users, and even 5M embedding tokens monthly on Pinecone-hosted models. For a prototype or an internal tool, that ceiling holds for months. There's also a 3-week trial with $300 in credits on the paid tiers, which is the right way to measure your own read-unit burn before committing.

The paid lineup has three rungs. Builder is new: a flat $20/month for solo developers and small teams that outgrew Starter but don't want a minimum commit. Standard carries a $50/month minimum spend — anything under that in usage still bills $50 — and beyond it, serverless pay-as-you-go rates apply: $0.33/GB/month for storage, $16–18 per million read units depending on cloud and region, and $4–4.50 per million write units. Egress is $0.10/GB with the first 100GB free monthly. Enterprise is a $500/month minimum with a 99.95% SLA, BYOC deployment into your own VPC, customer-managed encryption keys, and SCIM; HIPAA compliance is a $190/month add-on even on Standard.

The read unit is the line item that decides everything. A RAG chatbot answering 100,000 questions a month, with each query touching top-k retrieval plus a metadata filter, burns somewhere between 1 and 10 read units per query — call it 1–2M read units, or $16–36 at list rates. That's fine. The same app at 5 million queries with aggressive reranking and multi-query fan-out can hit 30M+ read units, which is $500+ a month before the storage line. My honest read after auditing two of these bills: Pinecone is priced to be cheap for prototypes and cheap for enterprises (where $500 is rounding error), and most expensive for the mid-market teams in between. It's also cloud-only at any price, which ends the conversation in some procurement reviews.

Qdrant: the open source vector database that keeps winning migrations

Qdrant is written in Rust, licensed Apache-2.0, and has become the default answer when teams search for a Pinecone alternative that doesn't meter queries. The open-source engine runs anywhere — Docker, Kubernetes, a $10 VPS — and the managed Qdrant Cloud free tier gives a single-node cluster with 0.5 vCPU, 1GB RAM, and 4GB disk, free forever, no credit card. It's small, but it runs the same API as production, with free cloud inference on selected models included.

Standard tier is usage-based on dedicated nodes across AWS, Azure, and GCP, managed through API, Terraform, Pulumi, or CLI, with a 99.9% SLA single-AZ and 99.95% multi-AZ, plus backups and disaster recovery. Premium adds a minimum spend, enterprise SSO, private VPC links, and 24/7 support. The interesting deployment option is Hybrid Cloud: Qdrant runs managed clusters on your own Kubernetes, so the control plane lives in Qdrant Cloud while the data never leaves your network — a genuinely useful middle ground for regulated teams that don't want to hire a platform engineer.

Two technical points matter for cost. First, Qdrant's quantization options — scalar, binary, and product quantization — compress vectors dramatically, and since memory is what you pay for on a node, compression is a direct discount. Second, there is no per-request meter anywhere. A million queries a day costs the same as ten. For retrieval-heavy products, that's the entire decision; for barely-used internal tools, a $50-metered competitor may actually be cheaper than an idle node.

Weaviate: the bill you can compute from your data shape

Weaviate pricing is the most legible in the category: you're billed by total stored vector dimensions plus storage, not by abstract units. The published Flex rates start at $0.00465 per million dimensions monthly and storage from $0.12/GiB, varying by index type, compression, cloud, and region. Run the napkin math: 1 million objects at 768 dimensions is 768M dimensions, which works out to roughly $4 a month of dimension charges at the cheapest published rate — before the Flex minimum and compression choices reshape the actual bill. The point isn't that Weaviate is cheap; it's that you can estimate cost from your dataset before creating an account, which no per-unit competitor offers.

Flex starts at $45/month minimum, pay-as-you-go, no commitment, on a shared cluster with replication, RBAC, and a 99.5% SLA. Premium is a prepaid contract from $400/month with dedicated deployment, up to 99.95% uptime, SSO/SAML, and — on Enterprise Cloud for AWS — HIPAA compliance. The always-free tier covers 100,000 objects, 1GB of memory, 10GB of disk, one collection with up to 3 tenants, and 2,000 embedding requests a day on Weaviate-hosted models. The Query Agent — natural-language queries compiled into filtered vector operations — includes 1,000 requests/month free, then $30 per organization monthly with 4,000 requests included.

One transparency note in Weaviate's favor: the company publishes changelogs for its price book, including a September 2026 update that re-rated backups and introduced RQ-4 compression pricing with an explicit explanation of why rates changed. Most database vendors reprice silently and let the invoice break the news. Weaviate is BSD-3-licensed and fully self-hostable, with the cloud adding the managed layer, the embedding service, and the agents on top.

Zilliz Cloud pricing: Milvus with an operations team behind it

Milvus is the heavyweight of the open source vector database world — designed for billion-vector indexes, with the GitHub stars to prove it. But self-hosting Milvus is a commitment: the architecture depends on etcd, object storage, and a message queue, which makes it the operationally heaviest engine in this comparison. Zilliz Cloud sells the managed version, priced around scale economics.

The free tier includes 5GB of storage, 2.5M vCUs monthly, and up to 5 collections — the most generous storage allowance of the four free tiers here. Standard adds serverless clusters from $0 and dedicated clusters with backups, encryption, and basic monitoring. Enterprise starts at $197/month dedicated with a 99.95% SLA, SSO, granular RBAC, audit logs, multi-replica scaling, and private endpoints; Business Critical adds CMEK, full-path encryption, and HIPAA eligibility for regulated workloads. BYOC deployment runs the whole thing in your cloud account.

The cluster-type pricing is where Zilliz differentiates. Performance-optimized clusters hold about 2 million 768-dimension vectors per compute unit, serve 500–1,500 QPS at roughly 10ms latency, and start at $63 per million vectors per month. Capacity-optimized clusters hold 8 million vectors per unit at 50–100ms and start at $16 per million vectors. Tiered-storage clusters hold 40 million vectors per unit at 100–1,000ms latency and start at $5 per million vectors — this is the price point nobody metered can touch. If your workload is a 500-million-chunk knowledge base queried by humans a few hundred times a day, tiered storage is a rounding error on your AWS bill; on Pinecone it's a budget line item with its own Slack channel.

Which vector databases fit which workload

Of the four vector databases above, here's how the decision actually shakes out in production:

• Prototype or side project: Qdrant's free cloud cluster or Weaviate's free tier. Both give a real managed cluster with the production API for $0 and upgrade without an engine swap.
• Retrieval-heavy RAG product with real traffic: Qdrant self-hosted or Zilliz dedicated. Zero per-query metering means your bill grows with data size, not with success. Pinecone's read units make viral weeks expensive.
• Mid-market SaaS that wants zero ops: Pinecone Builder at $20 flat, or Weaviate Flex at $45 minimum if you value cost predictability from day one.
• Cost forecasting is a procurement requirement: Weaviate, uncontested. Dimension-based billing is the only model you can compute from your dataset spec before signing anything.
• Billion-vector scale or tiered archival search: Zilliz tiered-storage at $5 per million vectors, or raw Milvus self-hosted if you have a platform team.
• Regulated industry with data-residency clauses: Qdrant Hybrid Cloud on your own Kubernetes, Zilliz BYOC, or Weaviate Enterprise Cloud. Pinecone's cloud-only posture and BYOC-on-Enterprise are workable but pricier paths.

The comparison table

Every figure below comes from the vendors' official pricing pages as of October 2026. This category reprices quietly, so treat these as quotes, not folklore — Weaviate alone publishes a price-book changelog.

Frequently Asked Questions

What is the best vector database for RAG in 2026?

For most teams: Qdrant if you can self-host or want unmetered queries, Weaviate if cost predictability matters more than anything, Pinecone if you want zero ops and your query volume is modest, Zilliz if you're past 100 million vectors. The honest ranking for a typical SaaS RAG app storing 1–10 million chunks and serving under 5 million queries a month: every one of the four handles the workload, so pick by billing model. Start on a free tier, measure your actual read-unit or vCU burn for two weeks, then commit to the vendor whose meter your traffic doesn't trip. Once retrieval is settled, the next cost center in a RAG stack is model routing — our LLM gateway comparison covers that layer — followed by quality measurement, which the LLM observability comparison handles.

Is Pinecone worth it, or are Pinecone alternatives good enough?

Pinecone is worth it in three cases: your team wants absolutely zero infrastructure work, you need dedicated read nodes for spiky hot traffic, or you're an enterprise where the $500 minimum and BYOC are procurement wins rather than costs. The Pinecone alternatives have closed the gap everywhere else — Qdrant matches the core search quality, Weaviate beats it on billing legibility, and Zilliz undercuts it badly at scale. After auditing two real Pinecone bills, my contrarian take is that the read-unit meter punishes exactly the products that succeed: the more your RAG app gets used, the more each query costs. Success taxes are a bad trade when unmetered alternatives exist.

Qdrant vs Pinecone: which costs less at production scale?

Qdrant, almost always, past the prototype stage. Pinecone's floor is $50/month and its read units at $16–18 per million mean a 30M-query month costs roughly $500+ before storage; Qdrant's cost is a fixed node that serves unlimited queries, and a node comfortably handling that load typically prices well below the equivalent read-unit spend. Below roughly 1M reads a month, the comparison flips — Pinecone's free Starter or $20 Builder undercuts paying for any node. The crossover point is around consistent traffic that burns more than $50–100 in monthly read units; past it, Qdrant's fixed-cost model compounds in your favor.

Can I run a vector database for free forever?

Yes, three ways. Self-host Qdrant or Milvus — both Apache-2.0 — on hardware you already pay for, which is genuinely free forever at the cost of owning uptime and upgrades. Use Weaviate's always-free tier (100K objects, one collection) or Qdrant Cloud's free single-node cluster, both permanent with no trial countdown. Or stay under Zilliz's free 5GB with collection discipline. The realistic ceiling of free infrastructure in 2026 is around 1–2 million 768-dimension vectors with light traffic — enough for most internal tools, docs-search features, and early-stage products.

How do I move embeddings from one vector store to another?

Easier than vendors admit, because the expensive part — your embeddings — is portable. Export vectors plus metadata and IDs (every engine here has an export or scroll API), reload in batches into the new store, and rebuild the index, which is compute time rather than engineering time. Two cautions: keep the same embedding model, because vectors from different models live in incompatible spaces and mixing them silently destroys recall; and dual-write for two weeks before cutting over, with shadow-read comparison on a sample of queries, so you measure recall drift instead of hoping for it. Budget a weekend for the migration script and a fortnight for confidence.

The Bottom Line

All four vector databases will store your embeddings and return neighbors in milliseconds — the primitive is solved, and the billing model is the product. Pinecone wins on polish and zero-ops elasticity, and its new $20 Builder tier finally gives small teams an honest entry point, even if the read-unit meter still taxes success. Qdrant wins the migration wars on pure economics: open source, unmetered queries, and quantization that turns memory costs into a knob. Weaviate wins on legibility — the only bill in the category you can forecast from your dataset — with a price-book changelog competitors should copy. Zilliz wins at scale, where $5 per million vectors on tiered storage makes every metered competitor look like a rounding error's evil twin.

My own setup after this comparison: Qdrant self-hosted on the product with 40M+ chunks and constant traffic, Weaviate Flex on the client whose finance team demands a cost figure before every quarter, and Pinecone Starter still serving the internal docs search it's never outgrown. Total vector-spend: one Kubernetes node and $45 a month. The retrieval layer of a RAG stack should never be the most expensive line item in it — and in 2026, it doesn't have to be.