Vector Database Pricing — The Definitive 2026 Comparison
Published 2026-04-10 · Reviewed 2026-09-12, corrections outstanding · 11 services compared · Figures compiled 2026-04-10 · 13 of 13 figures in the “Recent Pricing Changes” table come from our records for 1,493 developer tools · 10 pricing changes tracked
11
Services Compared
6
Generous Free Tier
2
Open Source (Free)
2
Limited / Trial
The vector database landscape in April 2026: 11 services across five categories — dedicated cloud vector DBs, open-source self-hosted, PostgreSQL + pgvector, embedded/local, and serverless. Vector databases are the fastest-growing infrastructure category in developer tooling, driven by RAG pipelines and AI applications.
Key trends: pgvector has democratized vector search — Supabase, Neon, and any PostgreSQL instance now support similarity search with zero extra infrastructure. Dedicated vector DBs (Pinecone, Qdrant, Weaviate) differentiate on scale, performance, and advanced features like hybrid search and reranking. The open-source ecosystem is strong: Qdrant, Milvus, Chroma, and LanceDB are all fully free to self-host. Serverless options (Upstash Vector, Turbopuffer) offer pay-per-use with no infrastructure management.
This guide covers: pricing tables, category breakdowns, vector storage and dimension analysis, self-hosted vs managed cost comparison, cost analysis for solo developers and teams, hidden costs, and best-for-use-case recommendations — compiled by hand from vendor pricing pages.
The pgvector advantage: If you are already using PostgreSQL (Supabase, Neon, or self-hosted), pgvector adds vector search with a single `CREATE EXTENSION vector` — no extra infrastructure, no extra cost, no extra API to learn. For production RAG at scale (1M+ vectors), dedicated vector databases offer better performance and more features, but pgvector is the fastest path from zero to working semantic search.
Category Breakdown
Vector databases fall into five categories, each optimized for different scale, deployment model, and integration needs.
A date in a heading below cites a record we hold, not a property of the vendor it names. These figures were compiled on 2026-04-10; a dated badge marks a card whose vendor we have recorded a change for since, and the date moves to the newer record on the day that record takes effect. Every record we hold for any vendor named below is in the full change log.
Free forever cluster: 1 GB storage on AWS/GCP/Azure. Supports filtering, payload indexing, quantization, multi-tenancy. Open-source Qdrant can also be self-hosted with no limits. Rust-based for high performance.
An unused Free Tier cluster is suspended after 1 week and deleted after 4 weeks of inactivity unless you reactivate it. (From qdrant.tech/documentation/cloud/create-cluster, read 2026-10-07.)
Weaviate Cloud offers a "Free Forever" cluster with 100,000 objects, 1 GB memory, and 10 GB disk. Weaviate is also open source and free to self-host.
A free Weaviate Cloud cluster is suspended after 7 days of inactivity, with its data kept, until you reactivate it. (From docs.weaviate.io/cloud/manage-clusters/create, read 2026-10-07.)
Free tier: 5 GB storage, 2.5M vector compute units/month, up to 5 collections. Based on Milvus (most popular open-source vector DB). Supports GPU-accelerated search, scalar filtering, hybrid search. Milvus open-source is also fully free to self-host.
Open-source AI-native embedding database. Self-hosted: fully free with no limits — runs in-process with Python or JavaScript, or as a standalone server. Chroma Cloud's Starter plan is $0 a month plus usage, with $5 in free credits; Team is $250 a month plus usage (trychroma.com/pricing, read 2026-10-07). Simple API: add, query, filter. Automatic embedding generation with pluggable models.
On Chroma Cloud, reaching a usage limit, whether yours or Chroma's, pauses the service until the limit is changed. (From trychroma.com/pricing, read 2026-10-07.)
pgvector extension included on all Supabase plans (including free). 500 MB database shared across all tables including vector data. HNSW and IVFFlat indexing. Works alongside Auth, Storage, Edge Functions, Realtime. Full SQL access for hybrid queries combining vector search with relational filters.
Supabase makes daily backups only for Pro, Team and Enterprise projects, so a free project has none, and Supabase recommends exporting it yourself with the CLI's db dump command. (From supabase.com/docs/guides/platform/backups, read 2026-10-05.)
A paused project can be restored from the dashboard for 1 year. After that, only its backup file and Storage objects can be downloaded. (From supabase.com/docs/guides/platform/upgrading, read 2026-10-05.)
Supabase notifies you when you exceed the Free Plan quota. Continued overuse brings restrictions that can include a read-only database, 402 responses to every API request, or paused projects. They lift when the quota refills next billing cycle or on upgrade to Pro. (From supabase.com/docs/guides/platform/billing-faq, read 2026-10-05.)
pgvector included on all Neon plans. 1 GB of storage per project (20 GB in total), up to 100 projects. Scales to zero after 5 minutes idle. Branching included (great for testing different index strategies). Paid plans are usage-based with no monthly minimum.
On the Free plan, scale to zero can't be disabled: computes suspend after 5 minutes of inactivity. (From neon.com/docs/introduction/plans, read 2026-10-04.)
Running out of CU-hours or network transfer suspends the compute until the next billing period or an upgrade. (From neon.com/docs/introduction/plans, read 2026-10-04.)
At the storage cap, inserts, updates and deletes fail until space is freed or the plan is upgraded. (From neon.com/docs/introduction/plans, read 2026-10-04.)
Neon's Azure East US 2, West US 3, and Germany West Central regions no longer accept new projects as of April 7, 2026; Free-plan projects there inactive for 90+ days may be deleted after October 5, 2026. (From neon.com/docs/import/azure-regions-deprecation, read 2026-10-07.)
Platforms that create databases for their users can apply for the Agent Plan. You must have a Neon Scale plan with a credit card on file and get Neon's approval. Neon sponsors the databases for your free-tier users. Paid usage gets up to $25,000 in credits and compute at $0.106 per compute-unit hour. (From neon.com/docs/introduction/agent-plan.md, read 2026-10-11.)
Neon gives $20 in free credits to try paid features, claimed through a promo link in the Neon Console. (From neon.com/pricing.md, read 2026-10-11.)
Neon's pricing page offers startups up to $100k in credits. The startup page describes the Databricks Startup Program, up to $200K in credits for Neon and Databricks for qualifying startups. (From neon.com/startups, read 2026-10-11.)
The Neon Open Source Program gives eligible open-source projects up to $5,000 a year in credits. The project must use a recognized open-source license and be self-hostable. (From neon.com/programs/open-source.md, read 2026-10-11.)
After the free plan, Neon's paid plans bill only for usage, with no monthly minimum: Launch charges $0.106 per CU-hour of compute and Scale $0.222, both charge $0.35 per GB-month of storage, extra branches cost $1.50 per branch-month, and network transfer beyond 500 GB per project costs $0.10 per GB. (From neon.com/pricing.md, read 2026-10-15.)
Neon Realtime has been in public beta since 2026-10-15. It provides live updates from SQL queries without polling and is free during the beta. Compute and egress are billed under the plan; beta runs in four AWS regions: us-east-1, us-east-2, eu-central-1 and ap-southeast-1. (From neon.com/docs/realtime/overview, read 2026-10-16.)
Atlas Vector Search included on M0 free clusters (512 MB, shared, always free on AWS/Azure/GCP). Supports approximate nearest neighbor search, pre-filtering, exact match. Integrates with MongoDB aggregation pipeline — combine vector search with document queries in a single query. Atlas Search (full-text) also included.
Embedded vector database built on Lance columnar format. OSS: fully free, runs locally or on your cloud storage (S3, GCS, Azure). Zero-copy reads, automatic versioning, multi-modal (text, images, video). LanceDB Cloud is closed to new accounts.
Serverless vector database: 200M vectors × dimensions, 1,536 dimensions, 10,000 queries/day. REST API — works in serverless/edge environments. Built-in embedding generation (no separate API call needed). Metadata filtering. No connection management required.
No free tier. S3-native architecture — vectors stored on S3 with warm cache layer. Every plan has a monthly minimum: $16 on Launch and $256 on Scale (turbopuffer.com/pricing, read 2026-10-07). Supports namespaces, metadata filtering, hybrid search.
Self-Hosted vs Managed Cost Comparison
Self-hosting can dramatically reduce costs but requires ops expertise. Here is the honest comparison.
Approach
Cost (1M vectors)
Cost (10M vectors)
Ops Burden
Best For
Qdrant self-hosted
$5–10/mo (VPS)
$20–50/mo
Medium (Docker)
Cost-sensitive teams with ops skills
Milvus self-hosted
$15–30/mo (VPS)
$50–100/mo
High (etcd, MinIO, Pulsar)
Large-scale with dedicated infra team
Chroma self-hosted
$5–10/mo (VPS)
$20–40/mo
Low (single binary)
Prototypes, small-to-medium RAG
pgvector (self-hosted PG)
$5–15/mo
$30–80/mo
Medium (PG admin)
Teams already running PostgreSQL
Pinecone (managed)
$0 (free tier)
$70–200/mo
None
Zero-ops requirement
Qdrant Cloud (managed)
$0 (free tier)
$30–100/mo
None
Easy start with OSS escape hatch
Zilliz Cloud (managed)
$0 (free tier)
$65–200/mo
None
Milvus users wanting managed service
The self-hosted trade-off: Self-hosting Qdrant or Chroma on a $10/mo VPS can handle 1–5M vectors comfortably. But you own backups, monitoring, security patches, and scaling. Managed services cost 2–5x more but eliminate ops entirely. The sweet spot: start on a managed free tier (Pinecone, Qdrant Cloud), move to self-hosted when costs matter and you have the ops capacity.
Cost Comparison by Team Size
What does each vector database cost for a solo developer versus a team building production RAG?
Service
Solo / Prototype /mo
Team / Production /mo
Key Cost Factor
Pinecone
$0
$70+ (Standard)
Usage-based (read/write units)
Qdrant Cloud
$0
$10–50
Per-cluster (RAM/disk)
Weaviate Cloud
$0 (Free Forever)
$45+ (Flex)
Per-cluster
Zilliz Cloud
$0
$65+
Compute units (CU)
Chroma
$0 (self-hosted)
$0 self-hosted; Cloud Team $250/mo + usage
Self-hosted: free; Cloud: usage-based
LanceDB
$0 (self-hosted)
$0–100+ (cloud)
Self-hosted: free; Cloud: usage-based
Upstash Vector
$0–5
$10–30
Per-query + per-vector
Turbopuffer
$16 (Launch minimum)
$16–50
Usage-based, with a monthly minimum
Supabase pgvector
$0
$25+
Per-project
Neon pgvector
$0
Usage-based, no minimum
Usage-based
MongoDB Atlas Vector Search
$0
$57+
Per-cluster
Hidden Costs
Vector database pricing has unique gotchas beyond the sticker price.
Embedding API Costs
The vector database is only half the cost. You also pay to generate embeddings. OpenAI text-embedding-3-small: $0.02/M tokens (~$0.01 per 10K documents). Cohere embed-v3: $0.10/M tokens. For 1M documents, embedding costs can exceed vector storage costs. Consider open-source embedding models (Sentence Transformers, Nomic) to eliminate this recurring cost.
Dimension vs Cost Trade-off
Higher-dimension embeddings (3,072 for OpenAI text-embedding-3-large) use ~4x more storage than lower-dimension ones (768 for many open-source models). This directly multiplies your storage costs. For most RAG applications, 768–1,536 dimensions provide sufficient quality at 2–4x lower cost. Use Matryoshka dimensionality reduction (supported by modern embedding models) to tune this trade-off.
Index Building Costs
HNSW indexes (used by most vector DBs) consume significant CPU and memory during construction. On free tiers with limited compute, indexing 100K+ vectors can take hours. Some services (Pinecone, Turbopuffer) handle this transparently; self-hosted solutions require you to provision sufficient RAM (typically 2–4x the vector data size for HNSW).
Query Latency at Scale
Free-tier performance varies dramatically. Pinecone Starter and Qdrant free clusters share resources — p99 latency can spike during peak hours. Serverless options (Turbopuffer, Upstash) have cold-start latency. pgvector on shared Postgres (Supabase free, Neon free) competes with other queries for CPU. Budget for dedicated compute if sub-100ms p99 latency matters.
Vendor Lock-in Risk
Proprietary vector databases (Pinecone, Turbopuffer) have no standard export format — migrating requires re-indexing all vectors. Open-source solutions (Qdrant, Milvus, Chroma) let you export and move freely. pgvector stores vectors in standard PostgreSQL — pg_dump includes your vector data. If portability matters, weight open-source or pgvector-based solutions.
Recent Pricing Changes
Vector database pricing is evolving rapidly as the market matures. See full change timeline for all tracked changes.
The free tier has changed. The current pricing page shows a Starter plan with $5 in free credits and usage-based pricing for writes, storage, and queries. Unconfirmed. We could not check this change against an archived copy of Chroma's page, so it does not count toward the rating.Source ↗
Turbopuffer cut the Launch plan's monthly minimum from $64 to $16. Its changelog dates the cut to June 2026, and the Internet Archive's copy of 2026-06-18 still shows $64. Scale's minimum stays $256 a month. Source ↗
Weaviate Cloud replaced its 14-day sandbox trial with an always-free cluster (100,000 objects, 1 GB memory, 10 GB disk, 1 collection), announced on 2026-06-17: "The Weaviate Cloud managed database is now genuinely free, without a credit card or time expiration." Source ↗
Neon moved its paid Launch and Scale plans to usage-based pricing with a $5 monthly minimum. The Free plan went from 0.5 GB of storage and 191.9 compute hours per account to 0.5 GB and 50 CU-hours per project, across up to 10 projects. Source ↗
Best-for-Use-Case Recommendations
Pick the Right Vector Database
Best for prototyping / getting started
Supabase pgvector or Neon pgvector — zero extra infrastructure, SQL queries, vectors live alongside your relational data. Go from zero to working semantic search in under an hour.
Best managed free tier
Pinecone (2 GB, ~1M vectors) or Zilliz Cloud (5 GB, about 1M 768-dimension vectors). Both offer generous free tiers with zero ops. Zilliz has more free storage; Pinecone has a larger ecosystem and simpler API.
Best for production RAG at scale
Qdrant (self-hosted or cloud) or Pinecone (fully managed). Qdrant offers the best price-to-performance for self-hosted; Pinecone for zero-ops. Both support hybrid search, metadata filtering, and handle 10M+ vectors reliably.
Best for self-hosted / maximum control
Qdrant (Rust, high performance, Docker-friendly) or Chroma (Python-native, simplest setup). For enterprise scale: Milvus (Zilliz-backed, GPU-accelerated). All three are fully open-source with no feature gating.
Best for serverless / edge
Upstash Vector (REST API, edge-native, free tier) or Turbopuffer (S3-backed, pay-per-use, no minimum). Both work in serverless environments where TCP connections are unavailable.
Best for multi-modal (text + images)
Weaviate — built-in vectorization modules for text (OpenAI, Cohere, Hugging Face) and images (CLIP). Generative search combines retrieval with LLM generation in a single query. Multi-tenancy built-in for SaaS.
Best for existing MongoDB users
MongoDB Atlas Vector Search — add vector search to your existing MongoDB cluster with no extra infrastructure. Combine $vectorSearch with aggregation pipeline for hybrid queries. Free on M0 clusters.
Frequently Asked Questions
What is a vector database and why do I need one?
A vector database stores high-dimensional numerical representations (embeddings) of data like text, images, or audio, and enables fast similarity search. You need one if you're building RAG (Retrieval-Augmented Generation) pipelines, semantic search, recommendation systems, or any AI application that needs to find similar items. Traditional databases can't efficiently search across hundreds of dimensions.
Should I use a dedicated vector database or pgvector?
For prototypes and small-to-medium workloads (under 1M vectors), pgvector in Supabase or Neon is the simplest choice — no extra infrastructure, SQL queries, and your vectors live alongside your relational data. For production RAG at scale (1M+ vectors), dedicated solutions like Pinecone, Qdrant, or Weaviate offer better query performance, more indexing options, and purpose-built features like hybrid search and reranking.
Which vector database has the best free tier?
For managed cloud: Pinecone offers 2 GB (~1M vectors) free and Zilliz Cloud (5 GB, about 1M 768-dimension vectors). For self-hosted: Qdrant, Chroma, Milvus, and LanceDB are all fully open-source with no limits. For the simplest setup: Supabase pgvector (500 MB, included with Postgres) or Neon pgvector (1 GB per project) require zero extra infrastructure.
What is the cheapest vector database for production RAG?
Self-hosted Qdrant or Milvus on a $5-10/month VPS is the absolute cheapest for small-to-medium scale. Turbopuffer bills usage with a $16 monthly minimum on its Launch plan. Qdrant Cloud ($10/month) and Upstash Vector (10K queries/day free) are also cost-effective. Supabase pgvector is free if your total database is under 500 MB.
How many vectors can I store in 1 GB?
It depends on dimensions. With 1,536 dimensions (OpenAI text-embedding-3-small): ~170K vectors per GB raw, but with indexing overhead expect ~100K-150K. With 768 dimensions (many open-source models): ~340K vectors per GB raw, ~200K-250K with indexes. With 3,072 dimensions (OpenAI text-embedding-3-large): ~85K vectors per GB. Lower-dimension models are more cost-effective for storage.
Data Source & Methodology
Powered by AgentDeals. The tables on this page were compiled by hand from official vendor pricing pages and have not been re-checked since. Pricing changes are tracked via our deal changes timeline (454 total changes tracked). The pricing changes we track are updated continuously; the tables above are not.
Query this data programmatically via our MCP tools or REST API — search for vector database services, compare vendors, or track pricing changes from your AI coding assistant.