Info
Pinecone is the fastest route to a working index, Qdrant the cheapest to run yourself, and Weaviate the best fit when hybrid keyword-plus-vector search matters. All seven products here were compared on published pricing, supported index types, official MCP server status, and API depth.
Pinecone is the best overall vector database software for teams putting retrieval into production right now, largely because it publishes per-read and per-write rates you can model before signing anything. Qdrant is the better call if you are self-hosting and care more about the compute bill than about operational convenience. Redis Essentials, starting at $5 a month, is the best vector database software for small business teams shipping a first AI feature on a small production index. Everything below is drawn from vendor pricing pages, official documentation, and vendor-owned GitHub repositories, checked in August 2026.
What Changed in Vector Databases This Year
Two things moved. The Model Context Protocol went from experiment to table stakes: every product here now ships a first-party MCP server, and Weaviate compiled one directly into the database binary. Pricing also got specific. A year ago most vendors answered "how much" with a contact form. Now Pinecone, Weaviate, Chroma, and Zilliz publish per-unit rates, which makes real cost comparison possible.
Why You Need a Vector Database
- Search that survives scale. Brute-force similarity over a NumPy array stops being viable around a million vectors, and the failure is a cliff rather than a slope.
- Filtered retrieval. Metadata filters let a single index serve many tenants, so you avoid running a separate store per customer.
- Recall you can tune. Graph indexes expose knobs like search window size, so you trade accuracy against latency deliberately.
- Durability and backup. Rebuilding an index from source documents after a node loss is slow and expensive.
- An agent-callable interface. An MCP server turns your index into something a model queries directly, with no bespoke retrieval service in between.
How We Evaluated
Each product was scored on four things: whether pricing is published in usable per-unit terms, index and filtering depth, AI and MCP maturity, and deployment freedom. Facts came from vendor pricing pages, official docs, and vendor-owned repositories only. The full scoring criteria live in our methodology.
1. Pinecone
Pinecone runs the index for you, and that is the entire pitch. There are no shards to size and no compaction jobs to watch. The trade is that you cannot run it anywhere except Pinecone.
Pricing: Starter is free. Builder is $20/month flat. Standard carries a $50/month minimum usage commitment, Enterprise $500/month. Storage is $0.33/GB/mo, write units are $4 to $4.50 per million and read units $16 to $18 per million on Standard, varying by cloud and region. Egress is $0.10/GB with 100 GB included monthly. A HIPAA add-on costs $190/mo on Standard.
Top Features
- Serverless indexes with no capacity planning
- Namespace isolation for multi-tenant applications
- Integrated embedding and document reranking
- Backup and restore against object storage
- Metadata filtering on every query
- HIPAA-eligible configuration as a paid add-on
Pros
- Published per-unit rates for reads, writes, and storage
- No cluster sizing, rebalancing, or compaction work
- Integrated inference removes a separate embedding hop
Cons
- No self-hosted or on-premises deployment at any price
- Its MCP server works only with integrated-embedding indexes
AI/MCP Integration: Official. Pinecone publishes a first-party server as the @pinecone-database/mcp npm package, documented with nine tools covering index management, upsert, cascading search across indexes, reranking, and documentation search. The docs note the restriction plainly: it supports only indexes with integrated embedding.
API Integration: Yes. A documented REST API with first-party client libraries, covering index management, upsert, and query.
Cloud Based: Yes, managed only. No self-hosted build exists.
Platforms: Managed cloud service, reached from any language through its REST API and client libraries.
Best For: Teams that want retrieval live this week and will pay to skip operations.
Editor score: 4.6/5. The clearest pricing on this list, held back only by the absence of a self-hosted path.
2. Qdrant
Qdrant is written in Rust, released under Apache 2.0, and can be run as a single container on a laptop or as a managed cluster.
Pricing: The free tier is a single-node cluster with 0.5 vCPU, 1 GB RAM, and 4 GB disk, free forever, including free cloud inference on selected models. Standard is usage-based, billed hourly against vCPU, memory, storage, and inference tokens, with no flat per-unit rate published on the pricing page as of August 2026. Premium requires a minimum spend that is not disclosed. Hybrid Cloud and Private Cloud are quoted on request.
Top Features
- Apache 2.0 license with no feature gating
- REST API with an OpenAPI 3.0 specification
- gRPC interface for latency-sensitive production traffic
- Payload filtering alongside vector similarity
- Hybrid Cloud deployment into your own Kubernetes
- Built-in cloud inference on selected models
Pros
- Genuinely free to self-host with no capability ceiling
- Same engine locally, in your cluster, and in their cloud
- Two official MCP tools that cover the common agent-memory pattern
Cons
- Cloud pricing is calculator-driven rather than published as rates
- Premium tier minimum spend is not disclosed publicly
AI/MCP Integration: Official. The qdrant/mcp-server-qdrant repository sits under the Qdrant organization and describes itself as an official Model Context Protocol server acting as a semantic memory layer. It exposes two tools, qdrant-store and qdrant-find.
API Integration: Yes. REST with a published OpenAPI 3.0 spec, plus a gRPC interface the project recommends for production-tier search.
Cloud Based: Yes, with managed, hybrid, and private options, and full self-hosting.
Platforms: Docker and Kubernetes for self-hosting, plus managed, hybrid, and private cloud.
Best For: Engineering teams that want to control where the data sits and what the compute costs.
Editor score: 4.5/5. Loses points purely because you cannot forecast a cloud bill from the pricing page.
3. Weaviate
Weaviate is the one to pick when your queries are half keyword and half semantics. It has offered BM25 alongside vector search for years, and in 2026 it did something none of the others did: it compiled an MCP server into the database itself.
Pricing: The free tier covers 100,000 objects, 1 GB memory, 10 GB disk, and one collection, with 2,000 embedding requests a day. Flex starts at $45/mo with vector dimensions from $0.00465 per million and storage from $0.12/GiB, at 99.5% uptime. Premium Shared starts at $400/month with dimensions from $0.003875 per million. Premium Dedicated also starts at $400/month, at 99.95% uptime, with dimensions from $0.002718 per million.
Top Features
- Hybrid BM25 and vector search in one query
- MCP server compiled into the database binary
- REST, gRPC, and GraphQL interfaces
- Multi-tenancy with per-tenant isolation
- RBAC governing agent read and write access
- BSD 3-Clause license for self-hosting
Pros
- Best hybrid search story of the seven
- Per-dimension pricing that drops as you commit
- Agent access governed by the same RBAC as everything else
Cons
- Three query interfaces means three sets of docs to learn
- The in-database MCP server is off by default and needs explicit enabling
AI/MCP Integration: Official. Weaviate's documentation states the MCP server is built into Weaviate but disabled by default for security, added in v1.38. It exposes four tools: weaviate-collections-get-config, weaviate-tenants-list, weaviate-query-hybrid, and weaviate-objects-upsert, with the write tool gated behind write access.
API Integration: Yes. REST, gRPC, and GraphQL, all documented as first-party interfaces.
Cloud Based: Yes, plus self-hosting under BSD 3-Clause and a bring-your-own-cloud option.
Platforms: Docker and Kubernetes for self-hosting, plus Weaviate Cloud and bring-your-own-cloud.
Best For: Retrieval where exact keyword matches and semantic similarity both have to count.
Editor score: 4.5/5. The hybrid search and the in-binary MCP server are both genuinely differentiated.
4. Zilliz Cloud and Milvus
Milvus is the heavyweight. It ships nine float index types, plus binary and sparse variants, which is more index surface than anything else here. Zilliz Cloud is the managed service from the team that builds it.
Pricing: Zilliz Cloud Serverless bills in virtual compute units at $4 per million vCUs. Inserting 1 KB of data costs 0.25 vCU, deleting an entity costs 1 vCU, and every read costs a minimum of 6 vCUs. Storage rates vary by region, cluster type, and project plan. Dedicated cluster rates are not published as flat figures on the public pricing page as of August 2026. Milvus itself is free under Apache 2.0.
Top Features
- FLAT, IVF_FLAT, IVF_PQ, IVF_SQ8 and SCANN indexes
- HNSW with SQ, PQ and PRQ quantization variants
- BIN_FLAT and BIN_IVF_FLAT for binary embeddings
- SPARSE_INVERTED_INDEX for sparse vectors
- Apache 2.0 license on the core engine
- Managed, dedicated and bring-your-own-cloud deployments
Pros
- The widest index selection of any product reviewed
- Sparse and binary embedding support out of the box
- vCU pricing is documented down to the operation
Cons
- Operating self-hosted Milvus at scale is a real job
- Dedicated cluster pricing needs a sales conversation
AI/MCP Integration: Official. The zilliztech/mcp-server-milvus repository is maintained by Zilliz and supports stdio, SSE, and streamable HTTP transports, with tools for search, query, and collection management.
API Integration: Yes. A documented REST interface plus first-party SDKs and a command-line client.
Cloud Based: Yes, with serverless, dedicated, and bring-your-own-cloud options, and full self-hosting.
Platforms: Docker and Kubernetes for self-hosted Milvus, plus Zilliz Cloud serverless and dedicated clusters.
Best For: Large corpora where index choice, quantization, and sparse vectors all matter.
Editor score: 4.4/5. Unmatched on index depth, marked down for operational weight.
5. Chroma
Chroma won its audience by being the thing you can pip install and have working in four lines. Chroma Cloud is the hosted version of that same engine, and its pricing is the most granular on this list.
Pricing: Starter is $0/month with $5 in free credits and usage-based billing beyond that, covering 10 databases and 10 team members. Team is $250/month plus usage with $100 in credits, 100 databases, and 30 team members. Usage rates are $2.50 per GiB written, $0.33/GiB per month stored, $0.0075 per TiB queried, and $0.09/GiB returned over the network. Enterprise is custom.
Top Features
- Local install to hosted with the same client code
- Python, JavaScript and Go client libraries
- Full-text search alongside vector similarity
- Metadata filtering on collections
- Per-operation billing across write, storage and query
- Apache 2.0 license on the core project
Pros
- The lowest-friction start of any product here
- Query billing at $0.0075 per TiB is cheap for read-heavy work
- Development and production run the same engine
Cons
- The $250/month Team tier is a steep step up from Starter
- Fewer index tuning controls than Milvus or Qdrant
AI/MCP Integration: Official. The chroma-core/chroma-mcp repository is maintained by the Chroma team and provides collection creation and retrieval through vector search, full-text search, and metadata filtering.
API Integration: Yes. A documented HTTP API with first-party Python, JavaScript, and Go clients.
Cloud Based: Yes, alongside local and self-hosted deployment under Apache 2.0.
Platforms: Local Python and JavaScript processes, Docker, and Chroma Cloud.
Best For: Prototypes and small production apps where developer speed beats tuning.
Editor score: 4.2/5. Excellent on ramp-up and read pricing, thinner on tuning controls.
6. MongoDB Atlas Vector Search
If your documents already live in MongoDB, adding a second database to search them is a decision worth questioning. Atlas Vector Search indexes embeddings inside the cluster you are already running, next to the documents they describe.
Pricing: The M0 tier is $0/hour, free forever, with 512 MB of storage. Flex is $0.011/hour, capped at $30/month, with 5 GB. Dedicated starts at $0.08/hour, or $56.94/month, scaling from 10 GB to 4 TB of storage and 2 to 96 vCPUs.
Top Features
- HNSW approximate search inside your existing cluster
- Exact nearest-neighbor search for exhaustive matching
- Vector embeddings up to 8,192 dimensions
- Vector quantization to cut index memory
- Aggregation pipeline filtering on the same documents
- Available on free, Flex and dedicated tiers
Pros
- No second datastore to operate, secure, or keep in sync
- Filters run against real document fields, not a copied metadata blob
- The free M0 tier is enough to prove the pattern out
Cons
- Requires MongoDB v6.0.11 or v7.0.2 and later for approximate search
- Index tuning surface is narrower than a dedicated vector engine
AI/MCP Integration: Official. The mongodb-js/mongodb-mcp-server repository is a first-party Model Context Protocol server for MongoDB databases and MongoDB Atlas, covering cluster creation, user management, and infrastructure inspection.
API Integration: Yes, through the official MongoDB drivers and the Atlas Admin API rather than a separate vector-specific endpoint.
Cloud Based: Yes on Atlas, with self-managed and local deployments supported through the Atlas CLI.
Platforms: Atlas clusters, self-managed MongoDB, and local deployments created with the Atlas CLI.
Best For: Teams already on MongoDB who want retrieval without a second system to run.
Editor score: 4.1/5. The right answer more often than dedicated-vector-store enthusiasm suggests.
7. Redis
Redis is the outlier here, and deliberately so. It is a low-latency data structure server that also indexes vectors well, which makes it a strong pick for semantic caching and agent memory.
Pricing: The free tier is $0, up to 30 MB. Essentials starts at $0.007/hour, listed as $5/month, covering 250 MB to 100 GB with a single database. Pro starts at $0.014/hour with a $200/month minimum, and the page notes the first $200 is free. Redis Open Source 8.0 and later is tri-licensed under RSALv2, SSPLv1, or AGPLv3.
Top Features
- FLAT, HNSW and SVS-VAMANA vector index types
- L2, inner product and cosine distance metrics
- KNN search and vector range queries
- Hybrid filtering across text, numeric, geo and tag fields
- Vector storage in either hashes or JSON documents
- Sub-millisecond reads on the same instance you cache with
Pros
- Three index types including SVS-VAMANA compression added in Redis 8.2
- Vector search on infrastructure many teams already operate
- Range queries answer "anything close enough" without guessing a k value
Cons
- The 30 MB free tier is too small to hold a meaningful index
- Pro carries a $200/month minimum that skips the middle of the market
AI/MCP Integration: Official. The redis/mcp-redis repository is maintained by Redis and described as a natural language interface for agentic applications, including query engine tools to manage vector indexes and run vector search.
API Integration: Yes, through the Redis wire protocol and official client libraries rather than HTTP, using FT.CREATE, FT.SEARCH, and FT.HYBRID.
Cloud Based: Yes on Redis Cloud, with self-hosting available under the Redis Open Source license.
Platforms: Redis Cloud, plus self-managed Linux, Docker, and Kubernetes deployments.
Best For: Semantic caching, agent memory, and any retrieval where latency is the hard constraint.
Editor score: 4.0/5. Excellent at its actual job, awkwardly priced between $5 and $200 a month.
Comparison Table
| Tool | Best For | Starting Price | Standout Feature | AI-MCP Support | API Integration |
|---|---|---|---|---|---|
| Pinecone | Managed simplicity | $20/mo flat | Published per-unit read and write rates | Official, 9 tools | REST |
| Qdrant | Self-hosted cost control | Free forever tier | Apache 2.0 with no feature gating | Official, 2 tools | REST + gRPC |
| Weaviate | Hybrid keyword and vector | $45/mo | MCP server inside the database binary | Official, 4 tools | REST + gRPC + GraphQL |
| Zilliz Cloud and Milvus | Index depth at scale | $4 per million vCUs | Nine float index types plus sparse | Official | REST + SDKs |
| Chroma | Fast developer ramp-up | $0 with $5 credits | Same engine local and hosted | Official | HTTP + SDKs |
| MongoDB Atlas Vector Search | Existing MongoDB users | Free M0 tier | Vectors beside the documents | Official | Drivers + Admin API |
| Redis | Low-latency agent memory | $5/mo | SVS-VAMANA compressed index | Official | Redis protocol |
Comparing vendors beyond these seven? The AI software directory lists profiles with pricing, deployment model, and review data side by side, and the data management category covers the storage layer.
How to Choose
- Settle managed versus self-hosted first. It constrains every later decision, and it is the only one that is expensive to reverse.
- Model your read volume, not your vector count. Pinecone charges $16 to $18 per million reads; at high query rates that dominates storage cost entirely.
- Match the index type to corpus size. Below a million vectors a flat index is exact and fast enough; above it, HNSW or a quantized variant is worth the tuning.
- Confirm the MCP server's scope, not just its existence. Two tools and nine tools are very different products, and Pinecone's requires integrated-embedding indexes.
- Count the systems you will operate. If you already run MongoDB or Redis well, adding a seventh service has a real cost that no pricing page shows.
What This Actually Costs
Take a small production workload: 5 million vectors at 768 dimensions, 2 GB of index storage, 3 million queries a month, and 5 million writes at load time.
On Pinecone Standard, storage is about $0.66 at $0.33/GB, the initial write load runs roughly $20 to $23, and 3 million read units cost $48 to $54. Read volume, not storage, is the bill. On Weaviate Flex, 5 million vectors at 768 dimensions is 3.84 billion dimensions, which at $0.00465 per million lands near $18, below the $45 monthly floor, so you pay the floor. On Zilliz Cloud Serverless, the same 5 GB write load is roughly 1.25 million vCUs, or about $5.
Self-hosted Qdrant or Milvus on a single mid-size cloud instance lands in the $40 to $80 range in compute, with your own time as the uncosted line item.
Final Thoughts
Pinecone is the pick if engineering hours are your scarcest resource and you can live inside someone else's cloud. Qdrant is the pick if they are not.
The interesting case is the third one. If you already run MongoDB or Redis competently, the honest answer to "which vector database" may be "none of them." Atlas Vector Search on a $56.94 dedicated cluster does real work.
Weaviate deserves a mention for shipping an MCP server inside the database rather than beside it. Small architectural decision, large operational consequence: one process, one auth path, one RBAC policy.
More retrieval and agent tooling coverage sits in our AI and automation posts, and adjacent tooling in the data analysis category.