Elasticsearch vs OpenSearch Vector Search (2026)

Quick verdict

Short answer

Choose Elasticsearch if you want Elastic's vector features such as semantic_text with managed inference, BBQ quantization and retrievers for RRF and linear hybrid ranking, and you are happy with Elastic Cloud (on AWS, Google Cloud or Azure) or an Elastic subscription for features like bbq_disk. Choose OpenSearch if a permissive Apache 2.0 licence with no paid feature tier matters, you want a choice of vector engines (Faiss, Lucene, JVector), or you are on AWS and want Amazon OpenSearch Service. Both handle BM25, filters and hybrid ranking in one engine; the decision is mostly licence, managed platform and which vector features you need.

How we know: This comparison is research-based: history, licences, governance, vector field options and engines, hybrid search methods, embedding integration, managed services and prices were checked on 7 November 2026 against Elastic and OpenSearch documentation, licensing and pricing pages, release notes, the Linux Foundation press release and AWS documentation and pricing. We have not run either engine, so nothing here is a measured result.

Elasticsearch is Elastic's distributed search and analytics engine built on Apache Lucene. Vector search uses the dense_vector field (up to 4,096 dimensions) with approximate kNN on HNSW-based index types, optional int8, int4 or BBQ quantization, exact brute-force kNN, and semantic_text fields that create embeddings through an inference endpoint. Elastic moved the code from Apache 2.0 to SSPL and the Elastic License with 7.11 in 2021, and in 2024 added AGPLv3 as a third option for the free portions. The latest version in the release notes on 7 November 2026 was 9.5.6.

OpenSearch began as an AWS-led fork of Elasticsearch 7.10.2 after that licence change and reached 1.0 in July 2021. Since 16 September 2024 it has been governed by the OpenSearch Software Foundation under the Linux Foundation, and it remains Apache 2.0, including its k-NN and neural search plugins. Vector search uses the knn_vector field with the faiss engine (default; HNSW and IVF), lucene (HNSW and flat), jvector, or the deprecated nmslib. Release 3.9.0 was the latest on GitHub on 7 November 2026.

Same roots, different paths. Concepts such as indices, shards, mappings and BM25 carry over between the two, but their vector APIs have diverged since 2021. Treat a move between them as a migration, not a drop-in swap.

Side by side

Aspect	Elasticsearch	OpenSearch
Licence	Free portions under AGPLv3, SSPL 1.0 or ELv2 (your choice); paid features under Elastic subscriptions	Apache 2.0 for the engine and plugins; no paid feature tier
Governance	Developed by Elastic N.V.	OpenSearch Software Foundation (Linux Foundation) since September 2024
Vector field	dense_vector (float, bfloat16, byte, bit), up to 4,096 dimensions; sparse vectors	knn_vector from the k-NN plugin
ANN index options	hnsw, int8_hnsw, int4_hnsw, bbq_hnsw, flat variants, bbq_disk (Enterprise subscription)	Faiss (HNSW, IVF), Lucene (HNSW, flat), JVector; NMSLIB deprecated
Memory reduction	int8, int4 and BBQ quantization; DiskBBQ	Vector quantization, disk-based and memory-optimized vector search
Hybrid search	RRF retriever, linear retriever and ES\|QL combining lexical and semantic matches	hybrid query with a normalization processor (score-based) or score ranker processor (RRF, since 2.19); hybrid query since 2.11
Embedding integration	semantic_text with Elastic Inference Service, ELSER or third-party inference endpoints	Ingest pipelines with registered ML models or connectors to externally hosted models
Filtering	Filters inside approximate and exact kNN queries	Efficient k-NN filtering or post-filtering
Managed services	Elastic Cloud Hosted on AWS, Google Cloud and Azure; Elasticsearch Serverless	Amazon OpenSearch Service (domains and Serverless, up to OpenSearch 3.5); third parties such as Aiven
Managed pricing basis	Hosted from 99 USD per month (reference setup); Serverless per VCU-hour and GB retained	AWS: instance hours plus storage, no minimum; Serverless per OCU-hour plus storage
Main trade-off	Richer integrated vector and inference features, with some behind paid tiers and a more complex licence	Permissive licence and engine choice, with fewer managed choices outside AWS

Key differences

Licences and governance

Elasticsearch has changed licence twice. Up to 7.10 it was Apache 2.0; with 7.11 in 2021 Elastic moved to a choice of SSPL or the Elastic License; and in 2024 it added AGPLv3, which Elastic's licensing FAQ says was expected before 8.16 became generally available. Today the free portions are available under AGPLv3, SSPL 1.0 or ELv2 at your option, and paid features are under Elastic's commercial subscription terms. For internal use this is rarely a blocker, but redistribution or offering the engine as a service needs legal review.

OpenSearch is Apache 2.0 throughout, including the k-NN and neural search plugins, and has no paid feature tier in the engine. Governance moved from AWS to the OpenSearch Software Foundation, a Linux Foundation project launched with AWS, SAP and Uber as premier members. If a permissive licence and vendor-neutral governance are hard requirements, that settles the question in OpenSearch's favour.

Vector indexes and quantization

Elasticsearch keeps vector indexing inside Lucene and offers several HNSW-based index types with quantization: int8_hnsw, int4_hnsw and bbq_hnsw (Better Binary Quantization), plus flat variants for exact search and bbq_disk (DiskBBQ), which requires an Enterprise subscription. On 9.4 and later the default is bbq_disk where the licence allows it.

OpenSearch lets you choose the engine per field. Faiss is the default and supports HNSW and IVF (useful for very large datasets); Lucene supports HNSW and flat; JVector is also available; NMSLIB is deprecated, so older mappings that use it should be migrated. OpenSearch also documents quantization, disk-based vector search and memory-optimized search. Engine choice is more flexible, but feature behaviour can differ by engine, so check the engine matrix for the features you need.

// OpenSearch: knn_vector field on the Faiss engine with HNSW
PUT /products
{
  "settings": { "index.knn": true },
  "mappings": {
    "properties": {
      "embedding": {
        "type": "knn_vector",
        "dimension": 768,
        "method": { "name": "hnsw", "engine": "faiss", "space_type": "l2" }
      }
    }
  }
}

Hybrid search

Both engines combine BM25 and vector relevance in one request, which is their main advantage over a pure vector database. Elasticsearch uses retrievers: an rrf retriever fuses a lexical query and a semantic_text or kNN query by rank, and a linear retriever combines normalised scores with weights; ES|QL can also mix keyword and semantic matching. Elastic's subscriptions page lists RRF for hybrid search and the linear retriever, but the tier marks did not render as text, so confirm which subscription includes them for self-managed clusters.

OpenSearch uses a hybrid query (since 2.11) with a search pipeline. A normalization processor normalises and combines sub-query scores, and the score ranker processor, introduced in 2.19, applies reciprocal rank fusion instead. Both are part of the Apache 2.0 distribution.

Embeddings and inference

Elasticsearch has the more integrated path: a semantic_text field generates embeddings automatically through the Elastic Inference Service or a configured inference endpoint (Elastic-managed, third-party or self-managed models), and ELSER provides learned sparse retrieval. OpenSearch generates embeddings in ingest pipelines that call registered ML models, including connectors to externally hosted models. Both let you bring your own vectors instead.

Managed services

Elastic runs Elastic Cloud Hosted (resource-based, on AWS, Google Cloud and Azure) and Elasticsearch Serverless (usage-based by VCU-hour and GB retained). OpenSearch is offered mainly as Amazon OpenSearch Service, with managed domains billed by instance hours and storage, and OpenSearch Serverless billed by OpenSearch Compute Units; AWS supports OpenSearch versions up to 3.5 and legacy Elasticsearch up to 7.10. Other providers such as Aiven also run OpenSearch. In our view, teams on AWS that want deep integration with S3, Kinesis and other AWS services lean towards OpenSearch, while teams on Google Cloud or Azure, or that want Elastic's newest vector features managed, lean towards Elastic Cloud.

Pricing and licensing

Elasticsearch. Self-managed use of the free features costs nothing; paid self-managed subscriptions (Platinum, Enterprise) are quoted by node count and RAM. Listed on elastic.co on 7 November 2026 (USD): Elastic Cloud Hosted from 99 USD per month (Standard), 114 USD (Gold), 131 USD (Platinum) and 184 USD (Enterprise) for a reference configuration of 120 GB storage across 2 zones. Elasticsearch Serverless from 0.14 USD per ingest VCU-hour, 0.09 USD per search VCU-hour, 0.07 USD per machine learning VCU-hour and 0.047 USD per GB retained per month. A free trial is offered.

OpenSearch is free under Apache 2.0. Amazon OpenSearch Service (aws.amazon.com/opensearch-service/pricing, 7 November 2026, AWS examples for US East, USD): managed domains billed per instance hour plus storage with no minimum fee, for example 0.335 USD per hour for r6g.xlarge.search and 0.122 USD per GB-month for gp3 storage; OpenSearch Serverless at 0.24 USD per OCU-hour plus 0.02 USD per GB-month of storage. Rates vary by region and instance type. Third-party managed OpenSearch, such as Aiven, is priced by that provider.

Pricing checked on the vendors' official pages on 7 November 2026. Prices change; confirm before buying.

Where each one leads

Elasticsearch strengths

semantic_text fields with managed inference make embedding generation part of indexing
Several quantization options, including BBQ and DiskBBQ, to cut vector memory
Retrievers for RRF and linear hybrid ranking, plus ES|QL
First-party managed service on AWS, Google Cloud and Azure (Cloud Hosted), plus Serverless
ELSER for learned sparse retrieval

OpenSearch strengths

Apache 2.0 for the engine and plugins, with no paid feature tier
Vendor-neutral governance under the Linux Foundation
Choice of vector engine per field: Faiss (HNSW, IVF), Lucene or JVector
Hybrid query with score normalization or RRF in the open source distribution
Amazon OpenSearch Service and Serverless for teams on AWS

Limitations

Elasticsearch limitations

Some vector features, including bbq_disk, need an Enterprise subscription
Licence history (Apache 2.0, then SSPL/ELv2, then AGPLv3 added) needs legal review for redistribution
The subscription tier for RRF and other retrievers could not be confirmed from the subscriptions page text
Self-managed clusters need shard, memory and node sizing

OpenSearch limitations

Engine options change over time (NMSLIB deprecated), so older mappings may need migration
Feature behaviour differs between engines; check the engine matrix
Amazon OpenSearch Service lags the latest releases (up to 3.5 while 3.9.0 was current)
AWS pricing has several dimensions (instances, storage, OCUs), which makes estimates harder

When to choose each

Choose Elasticsearch if

You already run the Elastic Stack for search, logs or observability
You want automatic embeddings through semantic_text and managed inference
You want BBQ or DiskBBQ quantization and can license the tier that includes it
You need a first-party managed service on Google Cloud or Azure as well as AWS
AGPLv3, SSPL or ELv2 terms are acceptable for your use

Choose OpenSearch if

A permissive Apache 2.0 licence and neutral governance are requirements
You are on AWS and want Amazon OpenSearch Service or OpenSearch Serverless
You want IVF on Faiss for very large vector sets, or a choice of engine per field
You want hybrid search and RRF without any paid feature tier
You run OpenSearch already for logs and want to add vector search

When neither is right

You only need vector similarity, not full-text search and analytics: see Milvus vs Qdrant and Pinecone vs Weaviate.
Your data is in PostgreSQL and a modest vector workload can live there: see pgvector vs Pinecone.
You need large-scale ranking with custom models and tensor expressions: see the Vespa profile.
You want an in-process library for a single application: see Chroma vs LanceDB.

Final recommendation

Bottom line

Elasticsearch is the more direct fit if you want the most integrated vector and inference features in a search engine, a first-party managed service on all three major clouds, and you accept Elastic's licence options and subscription tiers. OpenSearch is the more direct fit if Apache 2.0 and foundation governance matter, you are on AWS, or you want a choice of vector engines without a paid tier. Both are capable hybrid search engines; list the vector features you need, check which Elastic tier includes them, and price Elastic Cloud against Amazon OpenSearch Service for your data volume.

Frequently asked questions

Is OpenSearch just Elasticsearch under another name?

No. OpenSearch began as a fork of Elasticsearch 7.10.2 in 2021, so older concepts carry over, but the two have diverged. Their vector fields differ (dense_vector in Elasticsearch, knn_vector in OpenSearch), as do their hybrid search APIs.

Is Elasticsearch open source now?

Elastic's licensing FAQ says the free portions are available under AGPLv3, SSPL 1.0 or ELv2 at your option. AGPLv3, added in 2024, is OSI-approved. Paid features remain under Elastic's commercial terms. OpenSearch is Apache 2.0 throughout.

Do both support RRF hybrid search?

Yes. Elasticsearch has an RRF retriever and a linear retriever. OpenSearch has a hybrid query with a normalization processor or, since 2.19, a score ranker processor that applies reciprocal rank fusion.

Which vector engines does OpenSearch support?

The k-NN documentation lists faiss (the default, with HNSW and IVF), lucene (HNSW and flat), jvector and nmslib, which is deprecated.

Can I run Elasticsearch on AWS?

Yes. Elastic Cloud Hosted runs on AWS as well as Google Cloud and Azure, and Elastic also sells Elasticsearch Serverless. Amazon OpenSearch Service, the AWS-run service, offers OpenSearch and legacy Elasticsearch versions up to 7.10 only.

Sources

Checked November 2026.

How we research comparisons: our editorial method.