Managed Vector Database Pricing in 2026: Pinecone vs Qdrant Cloud vs Supabase pgvector vs AWS

Pricing checked: 14 October 2026. This guide compares Pinecone, Qdrant Cloud, Supabase with pgvector, and Amazon RDS for PostgreSQL with pgvector. Prices and plan limits can change, so use the official pricing pages linked in the article before making a purchase.

A local RAG prototype may work with a laptop and a small collection of documents. Production changes the economics: the dataset grows, queries become concurrent, indexes need memory, backups need storage, and somebody must operate the service. The lowest advertised entry price is not necessarily the lowest total cost.

For many small applications, PostgreSQL with pgvector is a sensible first option because application records, document metadata and embeddings can live together. A dedicated vector service can be better when its search features, scaling model or managed operations justify the extra service.

Quick comparison

These are different products and pricing models. A free tier is useful for testing, but may not include production availability. A plan fee may include credits, while a usage-based service may meter several separate components. Compare the same workload and availability target.

1. Calculate raw embedding size

RAG systems store chunks, not just source documents. One PDF may become many chunks, and each chunk can have text, identifiers, URLs, timestamps and metadata. For float32 embeddings, raw vector bytes are approximately number of vectors × dimensions × 4.

This arithmetic is for raw vector values only. It is not a provider bill or the full database size. It excludes text, metadata, row overhead, indexes, replicas, logs, temporary index builds and backups. Index overhead depends on the index type and configuration. Compression and half-precision formats can reduce storage, but should be evaluated for retrieval quality.

Estimate vector data, payload data and operational headroom separately. Load a representative dataset, create the intended index, measure the actual footprint and leave room for updates and reindexing. Do not buy a plan that only fits the theoretical raw vectors.

2. Pinecone: managed search with a minimum spend

Pinecone's official pricing page lists Starter as free, Builder at $20/month, Standard with a $50 monthly usage minimum, and Enterprise with a $500 monthly usage minimum. Standard usage above the minimum is billed according to the service's pricing model. The page also advertises a three-week Standard trial with $300 in credits; verify current eligibility and terms.

Official sources: Pinecone pricing and Pinecone database cost documentation.

Pinecone's rate card includes storage, read units, write units and other product usage. The official page lists storage at $0.33 per GB per month and Standard write-unit rates at $4–$4.50 per million; Enterprise write-unit rates are $6–$6.75 per million. Read-unit rates vary by plan, cloud and region. These are rate-card inputs, not a complete estimate: units consumed depend on the dataset and query pattern.

Pinecone is worth evaluating when you want a managed vector service, its features fit your search workload, and your team values reduced operational work. It deserves more scrutiny for a small chatbot with irregular traffic, where the Standard minimum may be larger than the workload itself. Starter or Builder may fit a prototype, but confirm plan limits and production needs before relying on one for a critical service.

For an estimate, record plan, region, vector count, metadata size, query rate, top-k, filters, update frequency and any inference usage. Use the official calculator and compare the estimate with actual usage during a pilot. Do not multiply query count by a guessed flat rate.

3. Qdrant Cloud: resource-based pricing

Qdrant Cloud lists a free cluster with 0.5 vCPU, 1 GB RAM and 4 GB disk, intended for testing and prototypes. Paid Standard clusters use dedicated resources and support scaling; Premium is aimed at organizations needing additional features and support. Qdrant describes paid pricing as resource-based.

Official sources: Qdrant Cloud pricing and Qdrant billing and payments.

Qdrant says charges depend on CPU, memory and disk usage and recommends its calculator for an estimate. Usage is billed monthly in arrears. Paying through a cloud marketplace does not automatically mean that the cluster runs in your own cloud account; review the deployment model and data-residency requirements.

RAM matters alongside disk. The engine must access index structures, process filters and handle concurrent requests. Two datasets with the same vector count can need different resources because of dimensions, payload size, index configuration, replication and query pattern. Use the free tier to validate integration, not to assume production capacity. Test latency and retrieval quality with realistic concurrency.

Qdrant's open-source deployment may reduce the vendor invoice for some workloads, but self-hosting transfers responsibility for patching, monitoring, backups, recovery and capacity planning to your team. Include engineering time when comparing self-hosted Qdrant with Qdrant Cloud.

4. Supabase with pgvector

Supabase supports pgvector, a PostgreSQL extension for storing embeddings and performing vector similarity search. The architectural advantage is simplicity: user records, document metadata, application data and embeddings can live in one managed Postgres database.

Official sources: Supabase pricing, Supabase pgvector documentation and Supabase vector-index guidance.

Figures checked on 14 October 2026 describe the published rate card, not a promise that every RAG workload fits Micro. Additional usage, storage, egress, project count and optional features can change the bill. Read the billing page and monitor usage.

When an application already needs Postgres, pgvector can avoid a second database invoice and simplify joins with relational data. Supabase documents HNSW and IVFFlat indexes; the correct choice depends on the workload and quality targets. Use dimensions that match your embedding model, stable document IDs and retrieval logic that applies the right tenant and permission filters. Keep privileged database access on the server, and test that results are restricted to the requesting user.

As traffic grows, you may need larger compute, more disk, replicas, point-in-time recovery or other paid features. RAG queries can compete with transactional queries on one database, and index builds consume resources. Compare the fully configured Supabase bill with a dedicated vector service rather than comparing only the $25 plan headline.

5. Amazon RDS for PostgreSQL with pgvector

Amazon RDS for PostgreSQL is worth considering when your application already runs on AWS or your team wants managed PostgreSQL within that architecture. PostgreSQL can use pgvector for embeddings and similarity search, subject to engine-version and extension availability. RDS is managed relational PostgreSQL, not a dedicated vector database.

Official sources: Amazon RDS for PostgreSQL pricing and pgvector project documentation.

AWS charges for components including DB instance hours, provisioned storage, backup storage, data transfer and certain I/O or provisioned-performance options. Cost depends on region, instance class, storage type, deployment mode and usage. Multi-AZ changes availability and price. Use the AWS Pricing Calculator with a specific region and instance instead of quoting one universal monthly rate.

RDS may suit teams already operating AWS networking, monitoring and backup policies. It may be excessive for a small project that only wants a managed vector index. Compare the whole architecture: database, application compute, transfer, monitoring, backups and engineering effort.

6. Sizing examples

This is a capacity-planning guide, not a price quote. Pinecone has usage meters and minimums; Qdrant Cloud prices cluster resources; Supabase and RDS combine compute with storage and other costs. A fixed price based only on vector count would be misleading.

7. Build a monthly cost model

Use the same assumptions for each provider: number of source documents and chunks per document; embedding dimensions; text and metadata size; average and peak queries; top-k and filters; update frequency; availability target; region; and backup requirements.

Record the base fee or usage minimum, compute, disk, reads and writes, egress, backups, replicas, inference and applicable taxes. Calculate both a quiet month and a busy month. Use each provider's calculator, then compare the estimate with actual usage during a pilot. Include engineering time and migration risk if self-hosting is being considered.

8. Reliability and data handling

• Tenant separation: apply authorization and tenant filters on every retrieval path and test the behavior with different test accounts.
• Backups: confirm whether database records, source documents, indexes and configuration are all covered.
• Restore tests: measure recovery and reindexing time; an untested backup is only an assumption.
• Monitoring: track latency, errors, memory, disk, index size and retrieval quality.
• Data residency: confirm available regions and contractual requirements before uploading sensitive data.
• Embedding changes: changing models or dimensions may require re-embedding and reindexing.

Managed services have different support and availability features by tier. Read the terms for the exact plan rather than assuming a free or entry-level plan includes high availability or a contractual uptime SLA.

9. Make migration easier

Changing vector stores may require moving embeddings, chunk text, metadata, filters, index settings and application logic. Record the embedding model and version so query vectors remain compatible. Keep original source documents and a reproducible ingestion pipeline. Store stable document and chunk IDs and version the chunking configuration. After migration, rerun a fixed retrieval evaluation set and compare quality, latency and cost under the same load.

10. Final recommendation

Start with Supabase + pgvector if you already use PostgreSQL and want one managed database for application data, permissions and embeddings. Pro starts at $25/month, with compute credits and separate resource limits; measure your database before assuming Micro is enough.

Choose Pinecone when its managed vector-database features justify the pricing model. The Standard plan's $50 monthly minimum matters for small workloads, and actual charges depend on usage and region.

Choose Qdrant Cloud when a dedicated vector engine is the better technical fit and its resource-based estimate matches your measured needs. Use the official calculator rather than guessing a price per million vectors.

Choose Amazon RDS + pgvector when managed PostgreSQL inside AWS is the right operational fit. Estimate instance, storage, backups, transfer and availability together.

For a small RAG app, the best value may be the database your team already operates well. For production, pay for the service that meets latency, reliability and recovery needs at a cost you can explain. A measured pilot is more useful than choosing based on vector count or a marketing headline.

Frequently asked questions

Is pgvector cheaper than Pinecone?

It can be when PostgreSQL is already part of the application and the vector workload fits its resources. Compute, storage, indexes, backups and engineering effort still count. Pinecone may justify its cost when managed operations are valuable.

Can I use a free vector database for production?

Free tiers are useful for prototypes, but check limits, regions, inactivity rules, backups, availability and support before relying on one for a critical service.

How much disk do one million embeddings need?

At 1,536 dimensions and float32, raw vector values are about 6.144 GB decimal. Metadata, chunk text, indexes, logs and backups add to that. This is a first sizing estimate, not a provider quote.

Does every RAG app need a dedicated vector database?

No. PostgreSQL with pgvector can be sufficient when relational filtering and transactions matter. A dedicated service becomes attractive when its features or scaling model solve a real problem.

How do I avoid surprise costs?

Use official calculators, monitor storage and query usage, estimate peak traffic and review invoices during a pilot. Include backups, replicas, transfer and embedding-generation charges rather than comparing only base prices.

Related Gyan Aangan guides

• How to Build a Private Local RAG System in 2026 — architecture and retrieval basics.
• Local RAG vs Long Context: Which Should You Use? — decide whether retrieval is needed.
• AnythingLLM + Ollama RAG Setup Guide — test a local document-chat workflow.
• Chat With Your Own PDFs and Notes Using LM Studio — start with a private prototype.

Pricing and product details checked on 14 October 2026. Figures are USD list prices unless stated otherwise, before applicable taxes. Provider rates, limits and features can change. Storage examples are theoretical raw-vector calculations, not provider quotes or performance benchmarks.