OpenSearch 3.6.0 vs Elasticsearch 9.4.1: The architectural divergence is now absolute

A technical comparison of the latest search engines, detailing how decoupled storage, licensing shifts, and native query languages have permanently split these former siblings.

For teams building cloud-native log analytics platforms where storage costs dominate, OpenSearch 3.6.0 is the correct choice due to its native decoupled storage-and-compute architecture. For teams building high-performance vector search applications or enterprise search with complex analytics, Elasticsearch 9.4.1 remains the superior engine due to its advanced ES|QL engine and tighter Lucene integration. Skip OpenSearch if you require native, low-latency machine learning inference directly inside the cluster. Skip Elasticsearch if your compliance team forbids non-Apache 2.0 dependencies.

Methodology

This comparative review analyzes OpenSearch version 3.6.0 and Elasticsearch version 9.4.1, based on technical documentation, release notes, and architectural specifications available as of May 2026. This analysis is prompted by a query from Reddit user Casq-qsaC_178_GAP073. Because direct, independent side-by-side performance benchmarks on identical hardware for these specific minor versions are not yet publicly archived, this review focuses on verified architectural differences, API compatibility, and licensing frameworks. We do not cover long-term cluster stability under extreme multi-tenant loads or edge-case query optimization for custom Lucene analyzers.

Storage architecture splits completely

The most significant architectural divergence in OpenSearch 3.6.0 is its commitment to decoupled storage and compute. OpenSearch now supports remote-backed indexes as a primary deployment pattern, allowing clusters to use cloud object storage like Amazon S3 or Google Cloud Storage as the single source of truth. This design allows users to scale down local NVMe storage significantly, using local disks primarily as a read cache.

In contrast, Elasticsearch 9.4.1 continues to rely on a local-disk-first architecture for active, writable indexes. While Elasticsearch offers searchable snapshots to query cold data stored on object storage, its primary indexing and search path for hot data remains tightly coupled to local high-performance storage. This difference fundamentally changes the economics of log retention, making OpenSearch substantially cheaper for long-term data storage.

Query languages go separate ways

The query interface is no longer unified. Elasticsearch 9.4.1 heavily promotes ES|QL (Elasticsearch Query Language), a piped query language that allows developers to filter, transform, and aggregate data in a single step. ES|QL executes on a new, highly parallelized execution engine designed to process queries across blocks of data rather than row-by-row.

OpenSearch 3.6.0 relies on its own Piped Processing Language (PPL) and a robust SQL plugin. While both engines still support the legacy Query DSL, the development experience has diverged. Elasticsearch's ES|QL is more tightly integrated into its native visualization tools, whereas OpenSearch's SQL and PPL interfaces are geared toward compatibility with broader open-source analytics tools.

Vector search and AI capabilities

Both platforms have invested heavily in vector database capabilities, but their implementations differ. Elasticsearch 9.4.1 features a native inference API that allows users to perform vector embeddings directly within the cluster using integrated models or external APIs. It also includes native support for advanced quantization techniques, reducing the memory footprint of high-dimensional vectors.

OpenSearch 3.6.0 approaches this through its Neural Search plugin and k-NN engine. It supports Lucene's HNSW implementation alongside NMSLIB and Faiss engines. While highly performant, the integration feels more modular and less cohesive than Elasticsearch's native machine learning pipeline.

What is interesting and what is marketing

The decoupled storage model in OpenSearch 3.6.0 is a genuine, non-incremental improvement. It addresses the primary pain point of running large-scale search clusters: the linear scaling of expensive local storage alongside data volume. By making object storage the source of truth, OpenSearch has evolved into a true cloud-native database.

Conversely, much of the "AI-native" branding on both sides is marketing copy. Both engines ultimately rely on Apache Lucene for their underlying vector search mechanics. The differences in vector search performance are largely driven by how each engine tunes Lucene's HNSW parameters rather than any proprietary algorithmic breakthroughs.

Additionally, Elastic's decision to add AGPLv3 as a licensing option alongside SSPL and ELv2 is a strategic positioning move to reclaim the "open source" label. However, for enterprise legal departments, the copyleft requirements of AGPLv3 still present compliance hurdles that the Apache 2.0-licensed OpenSearch 3.6.0 entirely avoids.

Pricing as of May 2026

OpenSearch 3.6.0 is distributed under the Apache 2.0 license, making it entirely free to run, modify, and distribute. Managed versions, such as Amazon OpenSearch Service, are billed based on standard cloud infrastructure consumption.

Elasticsearch 9.4.1 is available under a dual-license model (ELv2, SSPL, or AGPLv3) for self-managed deployments. The basic tier is free but lacks advanced features like cross-cluster search, machine learning inference, and advanced security. Commercial tiers (Platinum and Enterprise) require proprietary licenses with opaque, custom pricing negotiated directly with Elastic.

Verdict

The technical divergence between these two engines is now absolute. Choose OpenSearch 3.6.0 if you are managing massive log analytics pipelines where storage costs are the primary constraint, or if your organization mandates strict adherence to Apache 2.0 licensing. Choose Elasticsearch 9.4.1 if you are building search-as-a-service applications that require complex, real-time aggregations via ES|QL, or if you need a tightly integrated, low-latency vector search pipeline with native machine learning inference.

What we would test next

In a future benchmark, we would deploy both engines on identical AWS EC2 instances to measure indexing throughput and search latency on a 1TB dataset of mixed log and vector data. Specifically, we want to measure the latency penalty of OpenSearch's remote-store architecture against Elasticsearch's local-disk-first approach under heavy concurrent query loads.

The investor read

Search tooling spend is bifurcating. OpenSearch is capturing the commoditized, high-volume log analytics spend (displacing traditional Splunk and Elastic setups on cost), while Elasticsearch is fighting to capture the high-margin AI search and vector database spend. This makes Elastic's commercial business highly dependent on enterprise adoption of its proprietary features like ES|QL and native machine learning inference. For investors, OpenSearch's success signals a deflationary pressure on basic search and log infrastructure, forcing commercial vendors to build highly differentiated, proprietary application-layer features rather than relying on core database monetization.

Pull quote: “The technical divergence between these two engines is now absolute.”