Build1 publisher3 min readPublished
Pinecone charges about 72 read units for one query against a 10M-vector index
A September 2026 comparison pins the workload at 10 million 1536-dimension vectors and reports Qdrant between $250 and $947 a month, while Pinecone's serverless bill is set by the gigabytes each query scans.
The Engineer · Build desk
What happened
- A September 2026 comparison fixes one workload for all four engines: 10 million 1,536-dimension vectors, about 1 KB of metadata each, 100 to 300 QPS sustained, p99 under 100 ms and Recall@10 of 0.95.
- Its bottom line puts Qdrant first on overall balance, crediting it with the highest QPS and the lowest p99 at equal recall plus native hybrid search.
- Pinecone serverless bills 1 read unit per gigabyte scanned per query, and the article says a query at 10M by 1536 scans about 72 GB, so about 72 read units.
- The cross-cutting caveats treat recall as a setting you tune, since any engine can be pushed to Recall@10 of 1.0 by raising ef_search or nprobe at a cost in QPS.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- cost When the billing unit counts gigabytes scanned, a team that keeps ingesting pays more for traffic it was already serving. The charge lands on whoever grows the corpus.
- constraint A QPS ranking taken at equal recall is unusable by anyone who does not know the ef_search and nprobe settings behind it, so the comparison has to be re-run on local data before it can decide anything.
- capability Quantization lets a self-hosting team cut bytes scanned directly, and that counts for most under the billing model where the provider does the indexing.
- decision With Qdrant's managed band under its self-hosted band at both ends, a team choosing to run the binary itself at 10M scale is buying control and paying a premium for it.
Reconstruct the 72 GB from the workload spec and you can see what a query is being charged for. Ten million float32 vectors at 1536 dimensions is 61.44 GB; 10 million records carrying 1 KB of metadata add 10 GB; the total is 71.44 GB [1]. The scan is the whole collection, metadata included [1]. Under a unit priced per gigabyte, collection size sets the charge on each query [6].
Hold traffic flat, ingest more documents, and the queries you were already serving get more expensive [5]. At 100 QPS, the floor of the stated load, 72 RU per query is 7,200 RU per second [2]. Over a 30-day month that is roughly 18.7 billion read units, and at 300 QPS roughly 56 billion [3]. The article does not publish a per-RU rate, so those totals stop at units. Pinecone's Enterprise plan lists Dedicated Read Nodes and a BYOC option alongside a 99.95% SLA [13].
Bytes per record therefore set the price, and three of the four engines let you cut them: Qdrant lists scalar, binary and product quantization [9], pgvector has halfvec and binary quantization [10], Milvus carries IVF_SQ8 and IVF_PQ [11]. Binary quantization at one bit per dimension takes a 1536-dim vector from 6,144 bytes to 192 [4]. The scan does not drop 32-fold, because the metadata does not compress with it. Per record it falls from 7,168 bytes to 1,216, so 71.44 GB becomes about 12.16 GB, and 10 of those 12.16 GB are metadata [5].
The article's own caveats undercut its performance ranking. Qdrant is credited with the highest QPS and lowest p99 at equal recall [3]. Equal precision is the precondition, and meeting it takes four matches: the same 1536 dimensions, the same 1 KB per record, the same filter selectivity in the query mix, and the ef_search or nprobe values each engine used to land on 0.95 [7]. The article also warns that sustained ingestion, filter-heavy query mixes and cross-AZ network hops all degrade published numbers [8].
The cost bands are small money either way. Self-hosted at $388 to $947 a month is $4,656 to $11,364 a year, and managed at $250 to $450 is $3,000 to $5,400 [6]. Managed sits under self-hosted at both ends, 36% lower at the bottom and 53% lower at the top [7]. Published dollar figures cover Qdrant; Zilliz appears as a storage rate of $0.025 per GB per month [12]. All of it is USD list, collected on 2026-09-16, with a note that prices change frequently and should be verified before publishing [2].
The article's stated limit on pgvector at this scale is capacity: a materially lower ANN QPS ceiling at 10 million vectors, from post-filter semantics and memory-hungry HNSW [10]. If the index triples to 30 million vectors, each Pinecone query scans about 214 GB and bills about 214 RU [8]. Qdrant's figures are quoted as monthly totals, self-hosted or managed [4].
What to watch
- A published per-read-unit rate would turn the 18.7 billion RU per month figure into a dollar number that can be compared with Qdrant's bands.
- Disclosed ef_search and nprobe settings at Recall@10 = 0.95 would make the QPS and p99 ranking checkable.
- List price changes after 2026-09-16 would move the Qdrant cost bands the recommendation rests on.