Vector databases compared: pricing, sparse vectors and embedding
Vector search is not a separate purchase everywhere: with pgvector, Elastic, Redis and Atlas it arrives as a line on a bill you already carry. Our review priced 44 tiers on the vendors' own pages and ranks ten indexes by how much of the bill each one will name up front.
What does a vector database cost beyond the free tier?
Ten of ten vendors put a starting price on the page, cheapest $20 a month. That number decides almost nothing. A hobby project pays nothing almost everywhere: 9 vendors never charge for the small end. The middle of the paid range is $82. Each is a floor over a meter counting vectors stored and queries served.
Starting for free is the default. 9 vendors keep a zero-cost tier open indefinitely, and most of those are simply the engine you host yourself. Weaviate is the exception: its cloud sandbox expires by itself after fourteen days, and the standing free option there is the self-hosted build.
pgvector, Elasticsearch, Redis and Atlas do not price vector search separately. It rides the meters of the database around it. What matters for those four is what your existing cluster costs once an index sits on it.
Most paid entry tiers are monthly minimums, not allowances. The Flex plan at Weaviate and the Standard plan at Pinecone both charge their floor at zero traffic and bill usage on top of it.
Scale inverts the arithmetic. Databricks publishes two classes: an AI Search Standard unit holds 2M vectors, an AI Search Storage Optimized unit holds 64M. Zilliz sells dedicated clusters the same way, by class, where the rate per million vectors tracks the latency you accept.
We priced 44 plan rows here, and seventeen name no number at all. Those rows bunch at one end of every lineup: enterprise tiers, dedicated infrastructure, bring-your-own-cloud.
Ranked by a transparency score: pricing transparency 60%, user satisfaction 40%. Capability is not scored. It is the condition grid below. Prices are read from vendor pricing pages and re-checked per product on the dates shown. ComparEdge sells no vector databases and takes no payment for placement. How the score is built.
01 / 06
Vector databases ranked: open source, sparse vectors and built-in embed
Rows sit in score order, not price order. One logarithmic axis carries every vendor here, and the gap between two marks is real money. Two columns read pricing conduct. Three record capability: open source, sparse vectors and built-in embedding.
Sorted by transparency scorePriced tiers 27 / 44Full disclosure 2 / 10How to read this table
Reading the console
The axis carries monthly figures only, and these vendors are not metering the same thing. Pinecone, Weaviate and Chroma publish a monthly floor and bill usage above it. Atlas shows a monthly number that is an hourly rate multiplied by the hours in a month. Databricks sells compute per hour and never lands on a monthly axis at all. Its row reads as a rate card instead. Take every entry mark as the least a vendor will accept while the index is small. What no column can show is recall at your data size, or how much memory the index wants before any of this costs real money.
Price axis
One shared logarithmic axis, $20 to $500. A tick further right is genuinely dearer. The tall tick is the cheapest paid tier.
Diamond
A flat platform fee, billed per month, not per seat.
Granted
Listed on the vendor’s own pages on the verify date.
Conditional
Present, but some tiers carry no published price.
Not on the record
Absent from the recorded feature set. It means unlisted, not incapable.
04pgvectorFree, no paid plansFree tierAll tiers pricedOpen source listedSparse vectors not on the recordBuilt-in embed not on the recordFree76Alternatives to pgvector
05MongoDB AtlasFree tier1 of 3 tiers unpricedOpen source not on the recordSparse vectors not on the recordBuilt-in embed not on the record$57flat70Alternatives to MongoDB Atlas
06WeaviateFree tier2 of 5 tiers unpricedOpen source listedSparse vectors not on the recordBuilt-in embed listed$45flat68Alternatives to Weaviate
07ChromaFree tier1 of 3 tiers unpricedOpen source listedSparse vectors not on the recordBuilt-in embed listed$250flat64Alternatives to Chroma
08QdrantFree tier3 of 6 tiers unpricedOpen source listedSparse vectors listedBuilt-in embed not on the record$65flat60Alternatives to Qdrant
GrantedSome tiers sealedNot on the recordFlat account fee, not per seatScore ranks pricing transparency and user ratings, not capability. Capability is the grid.
Scroll the console sideways to reach the remaining conditions.
02 / 06
Every database's plans, verdicts and the date we checked each price
Each record carries the tiers we found and the date every price was checked. Rows are ordered by our score, and that score reads pricing conduct and user ratings, never query speed. A fast engine can rank in the middle here.
Embeddings that belong beside the Delta tables they came from, where the index syncs off the table instead of a pipeline someone has to maintain. Databricks sells no standalone edition of AI Search. The index makes sense only where the platform is already in place.
Critical gapThe architecture mandates native Delta Table ingestion for synchronization.
Plan table and expert take
Databricks AI Search: expert take
The AI Search Storage Optimized unit runs $1.28 an hour and holds 64M vectors, against $0.28 an hour and 2M on AI Search Standard. Databricks sells no standalone edition, so the meter runs inside a platform bill, at a rate that moves by region.
Where Databricks AI Search holds up
Seamless integration with Delta Lake and Unity Catalog
Auto-sync keeps vector index current without manual pipelines
Unified governance across data and vectors
No separate infrastructure for existing Databricks users
★★★★★4.7CE scoreG2 4.6 · 768 reviewsFounded 2013Verified July 8, 2026
Running retrieval in production without anyone on the team owning a cluster. Pinecone is managed only, with no self-hosted build, and a free Starter tier that holds up to five indexes in a single region.
Critical gapThe platform requires constant cloud connectivity, preventing local or air-gapped database deployment.
Plan table and expert take
Pinecone: expert take
The $50 Standard plan is a floor, not a ceiling, and the $500 Enterprise plan is another floor above it. Enterprise also charges more per million read and write units than Standard. The step up buys an SLA and compliance work, not a cheaper meter.
A search cluster that already answers keyword queries and now has to answer semantic ones from the same index. Elastic sells no separate vector product, so vectors ride the meters the deployment already pays.
Elastic shops add vectors without adding a line to the bill. There is no vector SKU: dense vectors ride the search, ingest and storage meters the deployment already pays, and vector profiles carry a 50 GB storage allowance before storage is charged at all.
Where Elasticsearch holds up
Combines vector search with world-class full-text search in one engine
1B+ downloads: vast operational expertise available
ELSER provides state-of-the-art sparse vector without custom models
A Postgres table that needs an embedding column and nothing else. pgvector is an MIT-licensed extension with HNSW and IVFFlat indexing. The software costs nothing, and the bill stays whatever your Postgres provider charges.
Critical gapLarge datasets require manual index tuning to maintain query performance.
Plan table and expert take
pgvector: expert take
Postgres shops buy nothing here. The extension is MIT-licensed and free, and the whole bill stays with the Postgres provider you already pay. The cost that arrives is operational: nothing in plain Postgres throttles an index build, and a big one competes with live queries.
Where pgvector holds up
No new infrastructure: runs inside existing PostgreSQL
SQL interface familiar to every developer
ACID transactions across vectors and relational data
Works on Supabase, RDS, Neon: all managed PG providers
★★★★☆3.9CE scoreG2 3.8 · 12 reviewsFounded 2021Verified July 8, 2026
Above this line, nearly every rung carries a number. Lower down, the enterprise and dedicated steps go quiet, and those are the steps a production index usually lands on.
Keeping embeddings in the same documents as the operational data they describe. Atlas keeps a 512 MB cluster running free forever, then steps up to dedicated instances that meter hourly instead of selling a monthly plan.
Critical gapUnpredictable costs arise during complex query scaling.
Plan table and expert take
MongoDB Atlas: expert take
Teams on Atlas are not buying a new product. Vector search has no rate card of its own: it runs on the same cluster compute as the database, arriving as a bigger instance rather than a new line. Egress is metered separately.
Where MongoDB Atlas holds up
Unified operational + vector database eliminates extra infrastructure
500k+ developers already familiar with MongoDB
Strong free tier and serverless option
Hybrid search combines vectors with full-text Lucene
Collections that need vectorization built into the database rather than an embedding pipeline standing beside it. Weaviate is BSD-3 licensed and costs nothing to run in your own cluster; its cloud sandbox is a fourteen-day trial that expires by itself.
Critical gapNetwork latency increases significantly when data centers lack regional proximity.
Plan table and expert take
Weaviate: expert take
The $45 Flex plan is a monthly minimum that charges at zero traffic. The meter under it is unusual: Weaviate bills the total dimensions of your vectors, not queries or rows. Its committed tier lists a lower rate per million dimensions, behind a higher floor.
Where Weaviate holds up
Open source with managed cloud option gives deployment flexibility
Built-in vectorizers reduce pipeline complexity
Knowledge graph cross-references unique in category
Active community and excellent documentation
★★★★★4.7CE scoreG2 4.6 · 29 reviewsFounded 2019Verified July 16, 2026
Prototyping retrieval locally, where the store runs inside the application process before anything is hosted. Chroma costs nothing on your own machine under an open licence, and the hosted side begins on a free Starter tier.
Plan table and expert take
Chroma: expert take
The $250 Team plan carries a hundred dollars of usage credit, and the credit does not roll into next month. Chroma's own calculator sizes a million written documents, six million stored and ten million queries near eighty dollars of usage, before the plan fee.
Where Chroma holds up
Runs in-memory, file-based, or client/server with zero setup
Python-native API offers an exceptionally simple developer experience
Deep, first-class integrations with LangChain and LlamaIndex
Built-in multi-modal API supports text, image, and audio embeddings
Official Docker images and Helm charts simplify deployment
★★★★☆4.3CE scoreG2 4.2 · 6 reviewsFounded 2022Verified July 16, 2026
Prototyping on a laptop now and moving the same engine to a managed cluster later. Qdrant is Apache 2.0, and running it on machines you own costs nothing at all, with cloud clusters metered hourly on CPU, memory and disk.
Critical gapThe current implementation of hybrid search functionality is technically cumbersome.
Plan table and expert take
Qdrant: expert take
Self-hosting teams pay nothing at all, at any size, under Apache 2.0. Cloud buyers get no published unit rates: Qdrant meters CPU, memory and disk by the hour and sends you to a calculator. The free cluster has no failover and no shard rebalancing.
Where Qdrant holds up
Top benchmark performance via Rust and quantization
Named vectors enable multimodal and complex search patterns
Binary quantization reduces memory 32x
Excellent documentation and developer experience
★★★★★4.6CE scoreG2 4.5 · 12 reviewsFounded 2021Verified July 8, 2026
Search across billions of vectors on an engine you can also run yourself. Zilliz Cloud is the managed build of open-source Milvus, sold in dedicated cluster classes rather than one flat tier.
Critical gapThe open source Milvus version demands high administrative maintenance.
Plan table and expert take
Zilliz Cloud: expert take
Billion-scale buyers meet the real spread. Zilliz sells dedicated clusters by class, and the low-latency class costs several times more per million vectors than tiered storage. Milvus also builds indexes on GPUs, and that capability follows into the managed cloud.
Where Zilliz Cloud holds up
CNCF project: battle-tested for billion-scale workloads
GPU acceleration and DiskANN for cost-efficient large-scale search
Distributed architecture with independent storage/compute scaling
Multi-vector search supports complex AI use cases
★★★★★4.5CE scoreG2 4.4 · 53 reviewsFounded 2017Verified July 8, 2026
A cache that already sits in front of the application and now has to serve nearest-neighbour queries too. Redis builds vector search into its cloud tiers, where the size of the dataset sets the plan you land on.
Plan table and expert take
Redis Vector Store: expert take
Redis Cloud Pro sets a $200 monthly floor on top of an hourly rate, and vector search is not what you pay for. The capability is built into the cloud tiers, so dataset size and plan move the bill, not the query count.
Where Redis Vector Store holds up
Sub-millisecond vector search latency for applications already using Redis
No new database to manage if Redis is already in your stack
HNSW index delivers high recall with low query latency at moderate scale
Compare any two vector databases: plans, limits and our score
vs
What the records say
Databricks AI Search publishes no monthly figure, so there is no team bill to line up against Pinecone.
Pinecone carries 2 of the 3 capability columns on the record; Databricks AI Search shows 1.
Users rate them level: 4.6 on G2 apiece.
Pinecone runs a free tier to start on; Databricks AI Search does not.
Databricks AI Search prices everything it sells; Pinecone leaves part of its lineup unpriced.
Pick Databricks AI Search for: Embeddings that belong beside the Delta tables they came from, where the index syncs off the table instead of a pipeline someone has to maintain.
Pick Pinecone for: Running retrieval in production without anyone on the team owning a cluster.
№ 01
Databricks AI Search
CE 92 · G2 4.6
Published plans, US$/mo
AI Search Standard$0.28 / hr
AI Search Storage Optimized$1.28 / hr
Team of 6Not published
Verified July 8, 2026
№ 02
Pinecone
CE 80 · G2 4.6
Published plans, US$/mo
StarterFree
Enterprise$500
BYOCContact sales
Team of 6$20 / mo
Verified July 16, 2026
Both price lists on the category axis
Databricks AI SearchUsage-priced, public rate card
Vector questions: a million vectors online, separate store or a setting
What does it cost to keep a million vectors online?
Less than most people expect, and the plan fee is the larger half. A million embeddings at 768 dimensions is a few gigabytes, which fits inside the free tier at several vendors here. Chroma's own calculator puts a million written documents, six million stored and ten million queries near eighty dollars of usage a month, before its plan fee. Storage is cheap. Query volume and data returned are what grow.
Is it cheaper to add a separate vector database or to switch on vectors in the database you already have?
If the parent database is running and the index is small, the mode almost always wins. Atlas has no separate rate for vector search: it runs on cluster compute you buy anyway. Elastic puts dense vectors on the same search and storage meters as keyword queries. The dedicated engines earn their price later, at high query volume or heavy filtered search, where a general database needs bigger instances to keep up.
When does the free tier stop being enough for a real workload?
The moment an outage starts to matter, not a row count. Every free cluster on this page is single-node and unreplicated. Qdrant's has no failover and no shard rebalancing; the free Zilliz cluster is shared and carries no SLA. Most teams pay the day the index goes in front of users, and that day usually arrives before the gigabytes run out.
Should I sign an annual commitment to bring the rate down?
Only once your usage stops moving. Zilliz says an annual commitment earns extra credits and does not publish the percentage. Weaviate lists a lower rate per million dimensions on its committed tier than on the flexible one. Qdrant sells its private cloud as a yearly contract through a marketplace. All three trade a discount for a forecast, and vector workloads are the kind that triple after one feature ships.
What does a tier actually meter: storage, queries or vector dimensions?
All three, in different combinations, and that is why two entry prices are rarely comparable. Weaviate counts the total dimensions of your vectors. Pinecone counts read and write units on top of storage. Chroma bills writes, storage, queries and egress on separate lines. Databricks and Atlas charge for the hours the index occupies compute. The plan price only sets a floor under whichever meter is running.
Is there a charge for the data a query hands back?
Often, on a line separate from storage and queries. Chroma prices network egress per gibibyte returned. Pinecone meters egress above a monthly allowance. Atlas bills data transfer apart from compute across nearly every part of the platform. The line stays invisible while you test with small result sets. It appears when an application starts returning hundreds of matches with full metadata attached.
Where does self-serve end on this shelf?
At the step most production deployments need. Private networking, customer-managed keys, an uptime SLA and compliance paperwork sit behind a conversation almost everywhere on this page, and so do the dedicated and bring-your-own-cloud deployments. You can price a prototype on your own. Pricing something regulated means talking to a salesperson first.
A vector mode costs less on paper, so what are you giving up?
Mostly tuning and headroom. The dedicated engines expose the index itself: quantization, sharding, sparse vectors for hybrid retrieval, and knobs that trade recall against latency. A vector mode inherits whatever the parent database gives it, and it competes with your operational traffic for the same memory. That trade is comfortable at a few million vectors and gets uncomfortable well before a billion.
Which of these can you run yourself for nothing?
Most of them, and the licence costs nothing however large the index gets. Qdrant is Apache 2.0, Weaviate is BSD-3, Milvus and Chroma carry open licences, and pgvector is an MIT extension for Postgres. What you take on instead is the infrastructure bill and the operations: backups, replication, upgrades, and the memory an index wants. Self-hosting moves the cost, it does not remove it.
What changes when the index reaches a billion vectors?
The rate per vector starts to outweigh the plan fee. Databricks prices an AI Search Standard unit at $0.28 an hour for 2M vectors and an AI Search Storage Optimized unit at $1.28 an hour for 64M. A billion vectors will not sit in memory at a sane price. Vendors sell classes that trade latency for capacity, and index builds move onto GPUs where the engine supports it.
Do these engines charge for the embeddings themselves?
Several do, on a meter separate from the database. Pinecone prices embedding and reranking calls per million tokens. Weaviate sells its embedding models per million tokens as an add-on to the plan. Elastic charges for its inference service the same way. The vendors without built-in embedding assume you arrive with vectors already made, in which case that cost sits with whoever generated them.
Field note 01
Where the free vector index ends and the production bill starts
Free here means real capacity, not a token allowance. A free Zilliz cluster carries five gigabytes across five collections. Atlas keeps 512 MB alive with no expiry date. Qdrant gives you a single node with four gigabytes of disk. 9 of these vendors do something similar, and none of those offers runs out on a clock. A prototype fits in all of them, which is exactly why the entry price settles so little.
The step out is not a plan change. Published entry figures start at $20, and the midpoint of the range sits at $82. At several vendors that figure is a monthly minimum rather than a ceiling. What you pay in production tracks how many vectors you keep and how hard you query them. Neither number appears on a pricing card, and both are measurable against your own data in an afternoon.
Field note 02
Sometimes the vector database is a setting on the database you already run
pgvector, Elasticsearch, Redis and Atlas never sell you a vector product. pgvector is an extension, and the whole bill belongs to your Postgres provider. Vector search in Atlas has no rate card of its own and runs on cluster compute you buy anyway. Elastic puts dense vectors on the same search, ingest and storage meters as keyword queries, with a 50 GB storage allowance on vector profiles. Redis builds the capability into its cloud tiers, where dataset size sets the price.
The question stops being which vendor to pick. It becomes what the increment costs on hardware already running, and the increment is usually memory. That is easier to measure against your own vectors than to look up on anybody's pricing page.
The verdict on vector databasesSigned review · Updated
Oleh KemFounder & Lead AnalystComparEdge Editorial
Almost nothing on this shelf costs money at the size you will test at. That makes the opening price the least useful figure on the page, and moves the decision to what production does to it.
Two questions are worth more than any comparison of tiers. Whether your existing engine covers this job without a second product, and whether the tier in front of you is an allowance or a minimum charged at zero traffic. Several vendors answer both in public and several answer neither. That gap decides most of the order below.
MethodWe read 44 plan rows across ten vendors, floors and usage meters alike, and dated every figure .
DisclosureCollection is tool-assisted; every verdict is written and signed by a human analyst.
05 / 06
Not sure which? Answer one, take a shortlist.Pick the line that sounds like your team. Each one opens the vendor built for it.
How this review is made. Prices are read from vendor pricing pages and re-checked on the dates shown against each product. Condition columns reflect the feature set recorded on the vendor’s own pages on that date. ComparEdge sells no vector databases and takes no vendor payment for placement. Where a vendor publishes nothing, this page says so rather than estimating. Ranking is by transparency score: pricing transparency 60%, user satisfaction 40%. What a product can do is shown in the condition columns and carries no weight in the number.
Oleh Kem
