Compare / Vector databases
Vector database pricing, normalized
For 100,000 vectors, Weaviate Free, Pinecone Starter, and Neon Free are $0 if you fit their caps (Weaviate 100,000 objects; Pinecone 2 GB / 1 million read units; Neon 1 GB and scale-to-zero). Qdrant Cloud Free (1 GB RAM / 4 GB disk, about 1 million 768-d vectors) also covers that size. At about 1 million 1,536-d vectors stored, Neon always-on Launch is about $21.85, Supabase Pro $25, Weaviate Flex $45 (the $45 minimum binds), and Pinecone Standard is about $117 for 1 million queries plus storage. Qdrant Standard and turbopuffer storage/write unit rates were not on the public HTML we fetched. Self-hosting on a Hetzner CX33 is $9.99/month if you will operate Qdrant, Milvus, or pgvector.
Prices from each provider's official pricing page, checked 18 Oct 2026. USD, excluding tax. Storage examples use Pinecone's published size for 1,536-d dense vectors with 1 KB metadata (1 million records = 7.15 GB; 10 million = 71.5 GB). 24/7 hours use 730 h. "Not published" means the page did not list that number. Estimates, not guarantees.
Price for the same stored vectors
Pinecone Details
| $20.00 Builder flat (10 GB / 2 million RU). Standard storage alone would be 7.15 × $0.33 = $2.36, but Builder is the lowest paid SKU that holds 1 million × 1,536-d. Query volume at this size usually exceeds Builder 2 million RU — see examples. | $0.33/GB Standard/Enterprise storage | $16 / million RU Standard $16–$18 per million read units by cloud/region; 1 RU per GB of namespace, min 0.25 RU/query. WU $4–$4.50 per million. | Starter $0; Builder $20 flat; Standard $50 min Starter: 2 GB, 2 million WU, 1 million RU. Builder extra usage is blocked, not billed. |
|---|---|---|---|---|
Qdrant Cloud Details
| — Standard vCPU/RAM/disk rates not on the public pricing page. Free (1 GB RAM) does not hold ~7.15 GB. | — Per-GB disk rate not published on qdrant.tech/pricing | Not published Billed hourly on vCPU, RAM, disk, backups, and paid inference tokens — unit $ not in the fetched HTML | Free $0; Standard usage-based, no published minimum Free: 0.5 vCPU / 1 GB RAM / 4 GB disk, ~1 million × 768-d. Suspends after 1 week idle; deleted after 4 weeks. |
Weaviate Cloud Details
| $45.00 Flex minimum. Raw dimension+storage for 1 million × 1,536-d is about $8.00, below the $45 floor. | $0.12/GiB Flex, from $0.12/GiB. Premium dedicated from $0.33/GiB. | Dimensions + storage Flex from $0.00465 per 1 million dimensions; queries bundled (no separate RU). Data transfer free for a promotional period. | Free $0; Flex $45 min Free: 100,000 objects, 1 GB RAM, 10 GB disk, 1 collection. Premium is contact-sales. |
Zilliz Cloud Details
| $199.29 1 Capacity CU always-on ($0.273/h × 730 h). Holds ~1 million × 1,536-d (Capacity is 8 million × 768-d per CU). | $0.025/GB Dedicated storage | Bundled in the CU No separate RU. Capacity-optimized ~100–300 QPS per CU (docs). | No free cluster; 1 CU minimum while running $200 credits are a promo, not a published free tier. Suspended clusters stop CU charges; storage still bills. |
turbopuffer Details
| $16.00 Launch minimum. Storage and write $/GB not in the fetched HTML, so this is a floor for ~7.15 GB, not a full bill. | — Storage $/GB-month not published on turbopuffer.com/pricing HTML | $1 / PB queried From Feb 2026 changelog: $1/PB, 1.28 GB min per query, with large-namespace discounts. Pinning is GB-hours (128 GB / 10 min floor). | Launch $16 min Scale/Enterprise higher; Enterprise contact ≥ $4,096/month. Launch usage above $16 is billed. |
Supabase (pgvector) Details
| $25.00 Pro including Micro. 7.15 GB fits in 8 GB included disk. Free is 500 MB — too small for this size. | $0.125/GB after 8 GB included (Pro, General Purpose) | SQL (no vector meter) Queries are Postgres. Compute SKUs: Micro $10 (in Pro credits), Small $15, Medium $60, Large $110. | Free $0; Pro $25 Free: 500 MB, pauses after 1 week idle. Pro from $25/month includes $10 compute credits. |
Neon (pgvector) Details
| $21.85 Always-on 0.25 CU Launch ($19.35) + 7.15 GB × $0.35. $0 on Free if the project can sleep and fit 1 GB — 7.15 GB does not. | $0.35/GB Launch/Scale | SQL (no vector meter) Compute $0.106/CU-hour Launch; 1 CU ≈ 4 GB RAM. Scale $0.222/CU-hour. | Free $0; paid has no monthly minimum Free: 1 GB/project, 100 CU-hours, scale-to-zero after 5 minutes (cannot disable). |
Self-hosted on Hetzner Details
| $9.99 CX33 (4 vCPU / 8 GB) for an in-memory 7.15 GB index. CX23 $6.49 is 4 GB RAM — too small if HNSW stays in RAM. | — Included on the VPS disk you pay for; no separate $/GB vector meter | Host CPU Qdrant, Milvus, or pgvector on the box. No vendor query meter. | Software $0 + VPS Qdrant Apache-2.0, Milvus Apache-2.0, pgvector PostgreSQL License. You operate backups and upgrades. |
What you actually get
| Pinecone | Pinecone serverless (dense / sparse / full-text) | Hosted. Starter AWS us-east-1 only; Builder+ AWS, Azure, GCP | 1 RU per 1 GB of namespace per query, min 0.25 RU | Official size: 1 million × 1,536-d + 1 KB metadata = 7.15 GB; 10 million = 71.5 GB | AWS, Azure, GCP on Builder+; Starter us-east-1 |
|---|---|---|---|---|---|
| Qdrant Cloud | Qdrant (HNSW). Same engine as open-source. | Hosted on AWS, GCP, Azure (Qdrant-owned infra). Hybrid Cloud attaches your infra. | No separate read-unit meter; you size the cluster | Free holds about 1 million 768-d vectors (official create-cluster docs), not 1 million 1,536-d | AWS, GCP, Azure regions on the Cloud pages |
| Weaviate Cloud | Weaviate (HNSW / HFresh). Same engine as open-source. | Hosted shared Flex; Premium shared or dedicated | Billed on stored vector dimensions, not per query | 100,000 objects exactly matches the Small example; not 1 million | Flex: limited AWS/GCP regions (7). Free: 2. Dedicated Premium: ~40 including Azure |
| Zilliz Cloud | Managed Milvus. Capacity vs Performance vs Tiered-storage CUs. | Hosted dedicated (AWS, GCP, Azure on paid dedicated) | CU-hour. Capacity CU ≈ 8 million 768-d vectors (official). | Performance CU ≈ 2 million 768-d; Capacity 8 million; Tiered 40 million | AWS, GCP, Azure (plan-dependent) |
| turbopuffer | Object-storage-backed ANN (turbopuffer) | Hosted multi-tenant; pinning reserved cache; Enterprise BYOC | TB queried (1.28 GB min/query) unless the namespace is pinned | No $0 tier published | Not listed as a region table on the pricing page we fetched |
| Supabase (pgvector) | Postgres + pgvector on Supabase | Hosted Postgres (Auth/Storage/Functions on the same bill) | Ordinary SQL; you size compute for HNSW RAM | 500 MB database — below 0.715 GB for 100,000 × 1,536-d + 1 KB metadata | Supabase regions |
| Neon (pgvector) | Serverless Postgres + pgvector | Hosted. Scale-to-zero on Free. | Ordinary SQL; you size CU for RAM | 1 GB — covers 100,000 × 1,536-d (~0.715 GB), not 1 million | See neon.com/pricing |
| Self-hosted on Hetzner | Qdrant / Milvus / pgvector on a Cloud VM | You run it. CX33 $9.99; CX43 $18.49 (8 vCPU / 16 GB) still cannot hold 71.5 GB in RAM. | Whatever the host CPU can serve | OSS is free; the VPS is not | Hetzner: Germany, Finland, US, Singapore |
Best for
- Pinecone
A managed dense index with published RU math, when $20 Builder or $50 Standard is acceptable.
- Qdrant Cloud
A $0 prototype on the 1 GB free cluster, then open-source Qdrant on a VPS if Standard rates stay unpublished.
- Weaviate Cloud
A $0 100,000-object cluster, then Flex at $45 when you outgrow Free and want a published dimension rate.
- Zilliz Cloud
Managed Milvus when you already want Zilliz and will keep a CU on. Expensive at 1 million vectors versus Flex or pgvector.
- turbopuffer
Object-storage-backed search with a $16 floor — only if you accept that storage/write rates were not in the HTML.
- Supabase (pgvector)
Vectors next to app data already on Supabase, when the $25 Pro floor is fine.
- Neon (pgvector)
pgvector that should sleep on Free, or a small always-on Launch compute at about $22 with 7 GB stored.
- Self-hosted on Hetzner
The cheapest 1 million vector production shape if you will run Qdrant, Milvus, or pgvector yourself.
A $0 free cluster vs a $45 Flex floor vs Postgres with pgvector vs a VPS you run
Hosted vector DBs meter different things. Pinecone bills storage, write units, and read units (1 RU per GB of namespace per query, minimum 0.25 RU). Weaviate Flex bills vector dimensions and GiB stored, with a $45/month minimum. Zilliz dedicated bills CU-hours ($0.273/CU-hour for Performance or Capacity) plus $0.025/GB-month. Qdrant Cloud Standard is usage-based on vCPU, RAM, and disk, but those unit rates are not on the public pricing HTML — omitted rather than guessed. turbopuffer publishes a $16 Launch minimum and query at $1/PB (1.28 GB minimum per query); storage and write $/GB were not in the fetched HTML.
pgvector on Supabase or Neon is SQL next to the rest of the database. There is no separate vector meter: you pay the Postgres plan plus storage. Neon Free is $0 with 1 GB and scale-to-zero after 5 minutes. Supabase Free is $0 with 500 MB and pauses after a week — too small for 100,000 × 1,536-d with 1 KB metadata (~0.715 GB). Always-on Supabase starts at Pro $25.
Self-hosting Qdrant, Milvus, or pgvector is $0 for the software plus a VPS. A Hetzner CX23 is $6.49/month (2 vCPU / 4 GB / 40 GB NVMe / 20 TB). That holds the Small example. Medium (7.15 GB) needs more RAM than a CX23 if the index is in memory — CX33 is $9.99 (8 GB). Large (71.5 GB) does not fit in CX43 RAM (16 GB) without quantization or on-disk indexes; the VPS price is still shown, with that caveat.
What would a real project cost?
Estimates, not guarantees. Built from the published rates above (checked 18 Oct 2026); check each provider's current pricing before you commit. Tax is not included. Numbers marked unverified or not published were left out rather than guessed.
- Small: 100,000 vectors, light RAG. 100,000 dense vectors × 1,536 dimensions, ~1 KB metadata (~0.715 GB), stored all month, 100,000 queries, negligible extra writes after the first load. Writes are omitted from hosted bills unless a plan quota would block them. Database may sleep.
- Medium: 1,000,000 vectors, production RAG. 1,000,000 × 1,536-d + 1 KB metadata = 7.15 GB (Pinecone's published example), stored all month, 1,000,000 queries. Initial ingest write-units omitted (one-time). Always-on.
- Large: 10,000,000 vectors. 10,000,000 × 1,536-d + 1 KB metadata = 71.5 GB, stored all month, 10,000,000 queries. Always-on. No dedicated-read-node or pinning SKUs (those are extra products).
| Pinecone | $0 Starter | $116.76 Standard storage + 7.15 million RU | $11,463.60 Standard storage + 715 million RU |
|---|---|---|---|
| Qdrant Cloud | $0 Free cluster | — Standard unit rates not published | — Standard unit rates not published |
| Weaviate Cloud | $0 Free; 100,000 objects | $45.00 Flex minimum | $80.00 Flex dimension + storage |
| Zilliz Cloud | $199.29 1 Capacity CU | $199.47 1 CU + storage | $599.66 3 Capacity CUs + storage |
| turbopuffer | $16.00 Launch minimum; storage rate unpublished | — Storage $/GB not published | — Storage $/GB not published |
| Supabase (pgvector) | $25.00 Pro; Free 500 MB too small | $25.00 Pro Micro; 7.15 GB inside 8 GB | $132.94 Pro + Large compute + extra disk |
| Neon (pgvector) | $0 Free; sleeps after 5 min | $21.85 0.25 CU Launch always-on | $102.41 1 CU Launch + storage |
| Self-hosted on Hetzner | $6.49 CX23 | $9.99 CX33 | $18.49 CX43; 16 GB RAM cannot hold 71.5 GB in memory |
Small: 100,000 vectors, light RAG
Try first: Weaviate Free at $0 — 100,000 objects is exactly the cap, 1 GB RAM / 10 GB disk. Pinecone Starter ($0, 2 GB / 1 million RU) and Neon Free ($0, 1 GB, sleeps after 5 minutes) also fit.
Runner-up: Qdrant Cloud Free at $0 (1 GB RAM is enough at this size) or Pinecone Starter if you want their SDK. Self-host on Hetzner CX23 at $6.49 if it must stay up and you will operate it.
Avoid: Zilliz at about $199 for a whole CU, Weaviate Flex $45, and Supabase Pro $25 when a $0 free cluster still holds 100,000 rows. turbopuffer's $16 Launch minimum is a floor, not a $0 tier.
Medium: 1,000,000 vectors, production RAG
Try first: Self-hosted Qdrant/Milvus/pgvector on a Hetzner CX33 at $9.99 if you will operate it. If you want hosted, Neon Launch always-on is about $21.85 (0.25 CU + 7.15 GB storage).
Runner-up: Supabase Pro at $25 (8 GB disk covers 7.15 GB) or Weaviate Flex at $45. Pinecone Builder is $20 but 7.15 million RU exceeds the 2 million RU cap (overage is blocked), so production query volume needs Standard at about $116.76.
Avoid: Zilliz at about $199 for 1 CU, and Qdrant Standard / turbopuffer storage — unit rates not published, so they are omitted rather than guessed.
Large: 10,000,000 vectors
Try first: Weaviate Flex at about $80.00 — dimension + storage rates, no per-query RU snowball. Self-host remains $18.49 on a Hetzner CX43 only if you quantize or use on-disk indexes; 16 GB RAM cannot hold 71.5 GB in memory.
Runner-up: Neon at about $102.41 (1 CU Launch + 71.5 GB storage) or Supabase at about $132.94 (Pro + Large compute + extra disk) if vectors should live in Postgres.
Avoid: Pinecone Standard at about $11,463.60, where almost all of the bill is 715 million read units. Zilliz at about $600 for 3 Capacity CUs is cheaper than Pinecone here but still far above Flex or pgvector.
Show the math
- Project — Small: 100,000 vectors, light RAG
- 100,000 dense vectors × 1,536 dimensions, ~1 KB metadata (~0.715 GB), stored all month, 100,000 queries, negligible extra writes after the first load. Writes are omitted from hosted bills unless a plan quota would block them. Database may sleep. Pinecone: 0.715 GB; 0.715 RU × 100,000 = 71,500 RU — inside Starter 1 million RU / 2 GB → $0. Qdrant Free $0. Weaviate Free $0. Zilliz 1 CU $199.29 + ~$0.02 storage. turbopuffer $16 min (query 100,000 × 1.28 GB × $1/PB ≈ $0.13). Supabase Free 500 MB too small → Pro $25. Neon Free $0 (0.715 GB < 1 GB). Hetzner CX23 $6.49.
- Project — Medium: 1,000,000 vectors, production RAG
- 1,000,000 × 1,536-d + 1 KB metadata = 7.15 GB (Pinecone's published example), stored all month, 1,000,000 queries. Initial ingest write-units omitted (one-time). Always-on. Pinecone Standard: 7.15 × $0.33 = $2.36 + 7.15 million RU × $16/M = $114.40 → $116.76 ($50 min does not bind). Builder cannot run 7.15 million RU. Qdrant paid: omitted. Weaviate Flex $45 min. Zilliz 1 CU $199.29 + $0.18 storage. turbopuffer: storage rate unpublished — omitted (Launch $16 is only a floor). Supabase Pro $25. Neon $19.35 + $2.50 = $21.85. Hetzner CX33 $9.99.
- Project — Large: 10,000,000 vectors
- 10,000,000 × 1,536-d + 1 KB metadata = 71.5 GB, stored all month, 10,000,000 queries. Always-on. No dedicated-read-node or pinning SKUs (those are extra products). Pinecone: 71.5 × $0.33 = $23.60 + 10 million × 71.5 RU × $16/M = $11,440 → $11,463.60. Qdrant paid: omitted. Weaviate: 15,360 million dims × $0.00465 = $71.42 + 71.5 × $0.12 = $8.58 → $80.00. Zilliz: 3 CUs × $199.29 = $597.87 + 71.5 × $0.025 = $1.79 → $599.66. turbopuffer storage unpublished — omitted. Supabase $25 + $110 − $10 + $7.94 = $132.94. Neon $77.38 + $25.03 = $102.41. Hetzner CX43 $18.49 (RAM caveat).
- Pinecone
- Official: Starter 2 GB / 2M WU / 1M RU. Builder $20 flat, 10 GB / 5M WU / 2M RU (overage blocked). Standard $50 min, storage $0.33/GB-month, RU $16–$18/M, WU $4–$4.50/M. Query = 1 RU/GB namespace, min 0.25. 1M×1536-d+1 KB = 7.15 GB (docs table). Source: https://www.pinecone.io/pricing/ (checked 18 Oct 2026)
- Qdrant Cloud
- Official Free: 0.5 vCPU / 1 GB / 4 GB, ~1 million 768-d. Standard is usage-based (vCPU, RAM, disk, backups, inference). Per-unit $ not listed on qdrant.tech/pricing or the billing docs we fetched — omitted. Source: https://qdrant.tech/pricing/ (checked 18 Oct 2026)
- Weaviate Cloud
- Flex min $45. Vector dimensions from $0.00465 / 1M. Storage from $0.12/GiB. 1M × 1,536-d = 1,536 million dims × $0.00465 = $7.14 + 7.15 GiB × $0.12 = $0.86 → $8.00, so $45 min. 10M: 15,360 × $0.00465 = $71.42 + 71.5 × $0.12 = $8.58 → $80.00. Source: https://weaviate.io/pricing (checked 18 Oct 2026)
- Zilliz Cloud
- List price Capacity/Performance $0.273/CU-hour → 730 h = $199.29 per CU. Storage $0.025/GB-month. 1 million 1,536-d ≈ half of 8 million 768-d → 1 CU. 10 million 1,536-d ≈ 15,360 million dims / (8 million × 768) = 2.5 → 3 CUs = $597.87 + 71.5 × $0.025 = $1.79 → $599.66. Source: https://zilliz.com/pricing/pricing-guide (checked 18 Oct 2026)
- turbopuffer
- Official Launch minimum $16/month (Jun 2026 changelog). Query $1/PB, 1.28 GB min/query (Feb 2026 changelog). Storage and write unit rates not in the pricing HTML — omitted. Source: https://turbopuffer.com/pricing (checked 18 Oct 2026)
- Supabase (pgvector)
- Pro $25 includes $10 compute credits (one Micro) and 8 GB disk. Extra disk $0.125/GB. 7.15 GB → $25. 71.5 GB → 63.5 extra × $0.125 = $7.94 plus Large compute $110 − $10 credits + $25 plan = $132.94. Source: https://supabase.com/pricing (checked 18 Oct 2026)
- Neon (pgvector)
- Launch compute $0.106/CU-hour. Always-on 0.25 CU = $19.35. 1 CU always-on = $77.38. Storage $0.35/GB-month. 7.15 GB × $0.35 = $2.50 → $21.85 with 0.25 CU. 71.5 GB × $0.35 = $25.03 + 1 CU $77.38 = $102.41. Source: https://neon.com/pricing (checked 18 Oct 2026)
- Self-hosted on Hetzner
- Hetzner EU USD excl. VAT after 15 Jun 2026: CX23 $6.49 (2/4/40 GB), CX33 $9.99 (4/8/80 GB), CX43 $18.49 (8/16/160 GB). Software $0. 71.5 GB in-memory needs more than 16 GB RAM — quantization or disk indexes, not a bigger SKU guessed here. Source: https://docs.hetzner.com/general/infrastructure-and-availability/price-adjustment/ (checked 18 Oct 2026)
FAQ
- What is the cheapest vector database for a small RAG app?
- Weaviate Free, Pinecone Starter, Qdrant Cloud Free, and Neon Free are $0 at 100,000 × 1,536-d. Supabase Free is 500 MB — too small — so Pro is $25. Zilliz starts at one CU (~$199).
- What does 1,000,000 vectors cost per month?
- Estimates: about $10 on a Hetzner CX33 if you self-host, $22 on Neon always-on, $25 on Supabase Pro, $45 on Weaviate Flex, about $117 on Pinecone Standard with 1,000,000 queries. Qdrant Standard and turbopuffer storage were not published as unit rates.
- Why is Pinecone so expensive at 10 million vectors?
- Queries bill 1 read unit per GB of namespace. 71.5 GB × 10,000,000 queries = 715 million RU × $16 per million ≈ $11,440, plus $23.60 storage. Weaviate bills stored dimensions once, not per query, so the same shape is about $80.
- Can I just use pgvector?
- Yes. Supabase Pro is $25 with 8 GB disk; Neon always-on Launch is about $22 at 7 GB stored, or $0 on Free if it can sleep and fit 1 GB. You size RAM for HNSW yourself. There is no separate vector query meter.
- Is Qdrant Cloud cheaper than Pinecone?
- The Free cluster is $0. Standard is usage-based, but vCPU/RAM/disk dollar rates were not on the public pages we fetched, so paid Qdrant is an em dash rather than a guess. Self-hosted Qdrant on Hetzner is $6.49–$18.49.
- What would a real project cost?
- Estimates from published rates: 100,000 vectors can be $0 on Weaviate, Pinecone, Qdrant, or Neon. 1,000,000 vectors and 1,000,000 queries are about $10 self-hosted, $22–$25 on Neon/Supabase, $45 on Weaviate, $117 on Pinecone. 10,000,000 vectors are about $80 on Weaviate Flex, $102 on Neon, $133 on Supabase, $600 on Zilliz, $11,464 on Pinecone.
Related pages
Stack of the Week
Get one stack a week that replaces a SaaS bill.