Vector Database pricing, compared

turbopufferLaunch

Assumes 1,536-dimension f32 vectors with no other attributes: 6,144 logical bytes (4 bytes per dimension), 6.144 GB per million, all in one namespace. Storage = $2.03 per million vectors per month. Writes assume batches of 3.1 MB or more (about 512 vectors) at the 50% discount: $6.14 per million vectors written; unbatched single-vector writes hit the 10 KB minimum ($20 per million). Queries: max(6.144 × vectors_m, 1.28) GB per query with the 32 GB/128 GB marginal discounts, at $1/PB — e.g. $6.14 per million queries at 1M vectors, $37.89 at 10M. Data returned ($0.05/GB) is left out — tiny for top-10 IDs and scores. f16 vectors halve storage and query bytes. Rates read from the pricing page's cost calculator; Launch and Scale share them (Scale adds SSO, HIPAA BAA and a private Slack). The minimum is applied to queries only, so the model overstates by up to the storage and write charges when queries cost less than the minimum. Bands stop at 100M vectors. Enterprise (from $4,096/month, 35% usage premium) is left out.

ChromaStarter

Assumes 1,536-dimension float32 vectors with no documents or metadata: 6,144 logical bytes, 5.72 GiB per million, each query searching one collection that holds all of them. Storage = $1.89 per million vectors per month; writes = $14.31 per million vectors written or updated; queries = 5.72 GiB × vectors_m ÷ 1,024 × $0.0075 per query, e.g. $41.91 per million queries at 1M vectors (less if your data is split across many collections and each query searches one). Data returned ($0.09/GiB) is left out — small when queries return IDs and distances only. Team's $100 is applied to queries only, so the model overstates by up to the storage and write charges when queries cost less than $100. New accounts get $5 of free credit (not modelled). Bands stop at 100M vectors. Sync, forking ($0.03 per fork) and Enterprise are left out.

QdrantStandard (3 nodes, 0.5 vCPU / 4 GiB each)

Qdrant publishes no per-unit rates; prices come from Qdrant Cloud's own capacity calculator (cloud.qdrant.io/calculator) for AWS us-east-1, 1,536 dimensions, no quantization, replication factor 1, vectors kept in RAM, at 24 hours × 30.5 days. The calculator sizes RAM at about 9.2 GiB per million such vectors (vectors plus index and payload headroom); each plan is its recommendation at the top of the band (0.1, 0.5, 1, 5, 10, 50, 100M vectors) — e.g. 3 × 4 GiB nodes for 1M. Between stops this overstates, since a smaller cluster may fit. Scalar or binary quantization, or keeping vectors on disk, cuts RAM several-fold and the price with it; replication for high availability multiplies it. The free cluster's 1 GB RAM matches the calculator's 1 GiB requirement for 100k vectors, so it's capped there. Queries and writes aren't priced (resource-based). Backups and paid inference are left out; other clouds and regions differ.

PineconeStandard

Assumes 1,536-dimension float32 vectors with an 8-byte ID and no metadata: 6,152 bytes per record, 6.152 GB per million (Pinecone's size formula, decimal GB), all in one namespace. Storage = $2.03 per million vectors per month. Queries: RU per query = 6.152 × vectors_m (min 0.25) at $16/M RU — e.g. $98.43 per million queries at 1M vectors. Writes: 1 WU per KB of request, so about 6.15 WU per vector in batched upserts (the 5-WU per-request minimum is negligible for batches), = $24.61 per million writes at $4/M WU; updates of existing records cost roughly double. Rates are the low end of Pinecone's ranges ($16–18/M RU, $4–4.50/M WU, varying by cloud and region). The $50 Standard minimum is applied to queries only, so the model is exact when queries alone cost at least $50 and otherwise overstates by up to the storage and write charges. Starter (free: 2 GB, 1M RU, 2M WU, 1 GB egress a month, us-east-1 only) and Builder ($20 flat: 10 GB, 2M RU, 5M WU, 10 GB egress) caps are converted to vectors and queries with the same per-record sizes. Not modelled: egress ($0.10/GB beyond 100 GB on Standard — small when queries return IDs and scores only), backups, imports, Inference and Assistant. Bands stop at 100M vectors. Enterprise ($500 minimum, 1.5× unit rates) is left out.

WeaviateFlex

Converted with the Flex formula in Weaviate's own pricing-page calculator, using its default Performance Optimized profile (HNSW index, RQ-8 compression) for 1,536-dimension vectors with no other properties: $252.59 per million vectors per month, of which about 93% is vector dimensions ($0.0424 per 1M dimensions × 3 replicas × 1.2) and the rest storage (15.4 GiB per million at $0.2125/GiB × 3) and 7 days of backups. The $45 minimum is exact here: it covers the first 0.178M vectors. The Cost Optimized profile (HFresh index) prices at $195.61 per million vectors and suits small or non-latency-sensitive workloads. The pricing table's 'from $0.00465 per 1M dimensions' is the cheapest index, compression and region combination, not this one. Rates are the calculator's base region; other regions cost more. Free: 100,000 objects, 1 GB memory, 10 GB disk, one collection, no replication. Premium (from $400/month, prepaid contract) is left out.