Updated October 24, 2026
· ~4 min read
Two launches changed this decision in the last year. Amazon S3 Vectors went generally available in December 2025 (vectors stored in S3 itself, up to two billion per index, paid per gigabyte and per query with no cluster at all) and in June 2026 AWS cut its query processing price for indexes over ten million vectors by up to 80%. Then in May 2026 the next generation of OpenSearch Serverless arrived, which scales to zero after ten idle minutes. The old reasons to avoid both (S3 Vectors was a preview, OpenSearch Serverless cost a few hundred dollars a month even when nobody used it) are gone.
Meanwhile Aurora PostgreSQL with pgvector kept getting better at the thing it was always good at: vectors next to the rows they describe, filtered with ordinary SQL. And Pinecone remains the managed specialist many teams start on.
This guide compares the four on what actually decides it: how much you store, how fast queries must be, whether you need keyword search alongside similarity, what an idle month costs, and how each plugs into Bedrock.
S3 Vectors in us-east-1 charges $0.06 per GB-month to store, $0.20 per GB to upload, $2.50 per million queries, plus a per-terabyte "data processed" charge per query that falls as the index grows — $0.004/TB for the first 100,000 vectors, $0.002/TB up to ten million, and $0.0004/TB beyond that since June 16, 2026. For a typical RAG corpus of a few million 1,024-dimension vectors, the monthly bill is usually tens of dollars rather than hundreds. OpenSearch Serverless charges $0.24 per OCU-hour for indexing and search. A classic production collection needs at least 2 OCUs — about $350 a month before storage, whether anyone searches or not — and a dev/test collection 1 OCU. Next-generation collections, now the default in the console, have no minimum: compute scales to zero after ten idle minutes and returns in about ten seconds, queueing requests rather than dropping them. Aurora's cost is the database you are probably already paying for; the questions are whether the HNSW index fits in memory on your instance class and whether vector queries will compete with your transactional load. Pinecone is priced by its own plans — check them against the volume you expect.
S3 Vectors is wrong when you need keyword relevance (product search, code search, anything where exact terms matter), more than 100 results per query, or consistent sub-50 ms latency on a user-facing path. OpenSearch Serverless is wrong when the corpus is huge and rarely queried — you will pay for capacity you don't use, even with the next generation, once traffic is steady. Aurora pgvector is wrong at hundreds of millions of vectors on one cluster, or when vector search would starve the OLTP workload sharing it. Pinecone is wrong when your security review requires the data to stay in your own AWS account, or when you would rather not add a vendor contract for something S3 now does.
Question 1 of 3
Do you need keyword and vector search together (hybrid), or aggregations?
Default to S3 Vectors for most RAG on AWS: it is the cheapest by a wide margin at any size, needs no capacity planning, and plugs straight into Bedrock Knowledge Bases — accept roughly 100 ms queries for frequently used indexes and up to a second for cold ones. Choose OpenSearch Serverless when you need hybrid search (keywords and vectors together), aggregations or consistently low latency under load; its next generation makes it affordable for spiky and idle workloads. Choose Aurora pgvector when the vectors belong to rows you already keep in Postgres and queries need joins, transactions or tenant filters you'd rather express in SQL. Choose Pinecone when you want a dedicated managed vector service that works the same across clouds. A common production pattern combines two: S3 Vectors for the long tail of cold content, OpenSearch for the hot, keyword-heavy set.
Production GenAI Platform on AWS
Deploy secure, scalable AI systems using AWS-native GenAI services. RAG, agent orchestration, and guardrails: production-ready.
Amazon Bedrock Enterprise AI Platform
A production AI platform on Amazon Bedrock: RAG pipelines, agent orchestration, guardrails, multi-tenant cost tracking, and the observability needed to run AI in production.
LLM Observability & Evaluation on AWS
See what your AI features actually do: every prompt, response, cost and latency traced, and evaluations that tell you whether a change made things better or worse. Running in your own AWS account.
We work with startups and growing companies on AWS architecture decisions. Tell us about your situation and we will recommend the right approach.
As of October 24, 2026. Provided as is, for general information, and accurate to the best of our knowledge on the date shown. Not legal, compliance, tax, financial or other professional advice. Prices, limits and vendor services change; check current vendor documentation before relying on anything here. Terms of use.