Cheapest Vector Database 2026: 8 Managed Stores Ranked

Verified as of 2026-10-18 against the official vendor pricing and documentation pages cited below.

The short answer: at 1 million vectors with the traffic profile below, Chroma Cloud is the cheapest managed vector database at roughly $5 per month, and it wins for one reason only, that it has no monthly minimum.^[1] The whole ranking turns on that single fact. Vendors that meter usage cheaply but enforce a floor (Pinecone, Weaviate, Turbopuffer) all cost more at solo scale because you pay the floor, not the meter. This pillar prices eight major managed stores against one fixed workload, shows the arithmetic, and links the per-vendor head-to-heads.

1. The workload we price every vendor against

One workload, applied to every vendor, so the comparison is honest:

• 1,000,000 vectors at 1,536 dimensions (the dimension of the common OpenAI text-embedding-3-small output), stored as float32.
• 10,000 queries per day (about 300,000 per month), a busy small product.
• 2,000 ingests per day (about 60,000 per month), a steadily updated corpus.
• 365-day retention, so the full index stays resident.

This is large enough to clear every free tier, which is the point: free-tier comparisons flatter whoever has the most generous trial, and tell you nothing about what you pay when the product works. At 1 million vectors you are paying real money everywhere, so the rate card is what decides the bill.

2. No-minimum vs floor: the only distinction that matters

Managed vector pricing comes in two main shapes, and the shape decides the winner more than the per-unit rate does. A third, Postgres priced by instance size, follows the memory the index needs.

No-minimum, pure usage. Chroma Cloud and Zilliz serverless bill only for what you store, write, and read, with no fixed floor.^[1][5] At small-to-mid scale these are the cheapest, because the bill is the metered usage and nothing else.

Floor-based. Pinecone Standard ($50/mo minimum), Weaviate Flex ($45/mo entry) and Turbopuffer Launch ($16/mo minimum) all carry a fixed monthly charge that applies even when your metered usage is lower.^[2][3][7] Turbopuffer's metered usage for our workload is about $5, but the $16 floor is what you pay. The floor, not the rate, sets the price at this scale.

Instance-sized. Supabase Pro is $25/mo per project with $10 of compute credit, enough for the 1 GB Micro instance, but pgvector's HNSW index answers fast only when it fits in memory.^[6] Supabase's own benchmark ran 1 million 1,536-dimension vectors on its 32 GB 2XL compute, $410 a month,^[9] so the price follows the index, not a floor.

3. The storage footprint, computed once

Storage footprint is the one figure that is identical across vendors, because it is set by the data, not the price list. The Embeddings DB Cost engine computes it from vector count and dimensions (4 bytes per float32 value, plus roughly 25 percent index overhead). The engine now models current rates with plan minimums, so its dollar columns print the same Pinecone $50 and Turbopuffer $16 minimums used in the ranking below. It only prices four of the eight vendors, so the full verified ranking is the table; the engine block is a consistent cross-check, not a replacement for it.

The engine reports a storage footprint of about 7.15 GB for this workload.^[8] That 7.15 GB is what every vendor's storage line is priced against in the table below. Its dollar columns now print plan-accurate figures (the Pinecone $50 and Turbopuffer $16 minimums match the ranking); they cover four of the eight vendors, so the full verified prices are still the numbers in the table.

4. The ranked comparison table

Ranked by the verified monthly cost for the 1-million-vector workload above. The cost column shows the arithmetic so you can re-derive it. Storage footprint is 7.15 GB throughout.

Rank	Vendor	Pricing model	Monthly cost at workload	How it is derived
1	Chroma Cloud	Usage, no minimum[1]	~$5	7.15 GiB x $0.33 storage + ~0.34 GiB x $2.50 writes + read charges, no floor
2	Zilliz serverless	Usage, no minimum[5]	~$0.29 storage + vCU compute	7.15 GB x $0.04/GB-mo storage + $4/M vCUs compute (vCU-per-query is calculator-gated)
3	Turbopuffer Launch	Usage + plan minimum[7]	$16	About $5 of usage (storage about $2 at $0.33/GB, writes under $1 at $2/GB, queries about $2 at $1 per PB queried), so the $16 Launch minimum applies
4	Weaviate Flex	Per-dimension + floor[3]	$45	1,536M dims x $0.00465 = ~$7 metered, but the $45 Flex minimum applies
5	Pinecone Standard	Per read/write/storage + floor[2]	$50	~$38 metered (about 7 read units per query on the 7.15 GB index, ~$34 of reads, plus writes and storage), so the $50 minimum applies; the ~2.15M read units a month exceed the $20 Builder plan's 2M
6	Supabase Pro (pgvector)	Plan + compute size[6]	~$426	$25 Pro + $410 2XL compute (the size Supabase benchmarks for 1M 1,536-dim vectors) - $10 credit + about $0.79 for the 14.3 GB of table and index above the 8 GB included disk[9]
n/a	Qdrant managed	Resource-based[4]	Not publicly documented	Per-vCPU-hour and per-GB-hour rates are calculator-only (free-forever tier exists)
n/a	Zilliz dedicated	Per CU, by cluster type[5]	Not priced for this workload	Published from $5 to $63 per million vectors a month by cluster type, but quoted for 768-dimension vectors; a 1,536-dimension index needs Zilliz's calculator
n/a	LanceDB Cloud	Usage, contact-gated	Not publicly documented	Pricing page redirects to a contact form; OSS engine is free, Apache 2.0

Read the table as two groups. The first six are rankable on verified figures, and the order is set almost entirely by whether a floor applies, or for Supabase by the memory the index needs. The last three cannot be ranked on price for this workload as of October 2026: Qdrant and LanceDB Cloud publish no rate card, and Zilliz quotes its dedicated rates only for 768-dimension vectors. Ranking them would mean inventing numbers, which this comparison does not do.

5. The vendors you cannot price for this workload

Three managed offerings cannot be priced for this workload from public figures, and it matters for any honest ranking, so name them plainly:

• Qdrant Cloud. Bills on the CPU, memory, and disk a cluster reserves, but the per-vCPU-hour and per-GB-hour figures are not on the pricing page; Qdrant directs you to its calculator.^[4] The one public anchor is an illustrative $85 cluster, not a flat rate. The free-forever 0.5 vCPU / 1 GB RAM / 4 GB disk tier is real, but it will not hold 1 million 1,536-dimension vectors.
• Zilliz dedicated. Serverless storage is a published $0.04/GB-month, and dedicated clusters now list per-vector rates: from $63 per million vectors a month on performance-optimized clusters, $16 on capacity-optimized and $5 on tiered storage.^[5] Those figures are measured on 768-dimension vectors, so a 1,536-dimension index still needs the calculator.
• LanceDB Cloud. The open-source engine is free under Apache 2.0, but the LanceDB Cloud pricing page redirects to a contact form with no rate card. There is no public per-unit figure to rank.

For these three, the verified public facts are the billing model, any free tier and, for Zilliz, the 768-dimension reference rates. Treat any third-party "estimate" of their managed price with suspicion, because the vendors themselves do not publish one.

6. Which to pick

1. Pure vector search, want the lowest bill: Chroma Cloud. No minimum means the bill is the metered usage, roughly $5 at 1 million vectors, and it scales down to near zero for a prototype with the $5 credit.^[1]
2. Already on Supabase Postgres: pgvector in the project you already pay for. Your vectors sit with your app data, in the same database and transaction, but budget for the memory: Supabase sizes 1 million 1,536-dimension vectors on 2XL compute ($410/mo), and half-precision vectors or a smaller index bring that down.^[6][9]
3. Want a free-forever managed cluster for a smaller index: Qdrant. Its free tier is genuine, though you size the paid cluster in its calculator once you outgrow it.^[4]
4. Need a contractual SLA and per-dimension billing: Weaviate, accepting the $45 Flex floor (99.5% SLA) or $400 Premium (99.9%); Enterprise reaches 99.95%.^[3]
5. Bursty serverless traffic, value a mature ecosystem: Pinecone, budgeting for the $50 minimum once you outgrow the free Starter and $20 Builder caps, and for read units that grow with the index.^[2]

Re-verify every rate before committing. Vector pricing moved a lot in early 2026 (Zilliz cut storage by roughly 87% to $0.04/GB-month, and Pinecone's read rate is now $16 to $18 per million read units, with each query billed 1 read unit per GB of index).^[5][2] Model your own numbers in the Embeddings DB Cost engine and confirm against the official pages.

7. Per-vendor head-to-heads

Each comparison below prices a pair of these vendors in detail. Start here, then read the matchup closest to your shortlist:

• Pinecone vs Qdrant pricing 2026: serverless per-usage vs free-forever cluster.
• Weaviate vs Qdrant pricing 2026: per-dimension billing vs per-cluster resource billing.
• Chroma vs LanceDB cost 2026: managed no-minimum vs embedded-on-object-storage.
• Milvus vs pgvector cost 2026: Zilliz managed Milvus vs Postgres-native vectors.
• Turbopuffer pricing 2026: object-storage plus operations, and the plan floor.
• Supabase pgvector free tier 2026: where the $0 tier holds and where it breaks.

Frequently asked questions

What is the cheapest vector database in 2026?

For a workload of 1 million vectors at 1,536 dimensions with 10,000 queries and 2,000 ingests per day, Chroma Cloud is the cheapest managed option at roughly $5 per month. It bills purely on usage with no monthly minimum: $0.33 per GiB-month storage, $2.50 per logical GiB written, and read charges of $0.0075 per TiB queried plus $0.09 per GiB returned. At this scale that lands near $5 per month. The vendors with a fixed floor cost more because the floor dominates: Turbopuffer Launch has a $16 minimum, Weaviate Flex starts at $45 and Pinecone Standard carries a $50 minimum (checked October 19, 2026). Supabase pgvector costs more again, about $426 a month, once its compute is sized to hold the index in memory.

Is Chroma Cloud really cheaper than Pinecone for a small RAG app?

At the 1-million-vector workload priced here, yes. Chroma Cloud has no monthly minimum, so its bill is the metered usage of roughly $5. Pinecone Standard meters lower per unit but enforces a $50 per month minimum usage commitment, so you pay the $50 floor, not the metered amount. Below the free ceilings the order can change: Pinecone Starter is free up to 2 GB, 1 million reads, and 2 million writes per month, and Chroma offers a one-time $5 credit (verified May 25, 2026).

Which vector databases do not publish their managed prices in 2026?

As of October 2026, three managed offerings cannot be priced for a 1,536-dimension workload from public figures: Qdrant Cloud (resource-based, calculator only), Zilliz Cloud dedicated clusters (per-vector rates published only for 768-dimension vectors), and LanceDB Cloud (the pricing page redirects to a contact form). You cannot rank these on verified public dollar figures, only on their published free tiers and billing models.

Is a self-hosted vector database cheaper than a managed one?

It can be on raw infrastructure cost, but not once operational time is counted. Self-hosting pgvector, Milvus, or the LanceDB open-source engine on your own object storage or a small Postgres instance can run for the price of the underlying compute and storage alone, which at 1 million vectors is single-digit dollars per month. The hidden cost is the engineering time to run, back up, and scale it. For a solo founder, a no-minimum managed store such as Chroma Cloud is usually the lower total cost.