Create a collection. Send data. Query. That is the entire setup for Amazon OpenSearch Serverless. No cluster sizing spreadsheet. No capacity planning meeting. No argument about how many nodes you need for peak versus average. You create, you use, you pay for what you used. When you're not using it, you pay for storage and nothing else.
That last part is new, and it is the part worth getting excited about.
The joy of just using it
OpenSearch Serverless scales compute to zero when there is no activity. Not "scales to a minimum baseline." Zero. Your dev/test collection that nobody touches on weekends costs you storage and nothing else until Monday morning. Your QA environment that runs a test suite for 20 minutes a day pays for 20 minutes of compute. Your production search that handles a thousand queries per second at noon and twelve queries per second at 3 AM pays for a thousand at noon and twelve at 3 AM.
Indexing and search compute scale independently. This is the part that deserves a pause. If you have steady-state ingest running (a CDC pipeline feeding documents from your database, say), the search side still scales down when nobody is querying. You do not pay search compute costs to keep your ingest running. Two workloads, two scaling curves, one collection. The economics of running both together just changed.
Autoscaling responds in seconds. When a marketing campaign drives a traffic spike at noon, compute scales up immediately to match demand. Your users see the same latency at 10x traffic as they saw at 1x. When the campaign ends and traffic subsides overnight, compute scales back down just as fast. By morning, if nobody is querying, the collection is back at zero compute cost. This cycle happens automatically, every day, without alarms, without manual intervention, and without you ever provisioning for a peak that may or may not arrive. The infrastructure mirrors your actual workload in real time, and the bill reflects what actually happened rather than what you feared might happen.
What this looks like in practice
In a recent demo, our team built a hybrid search application (semantic vector search combined with keyword search) against a dataset of 5,000 movies. The entire setup happened through OpenSearch Agent Skills: pre-built capabilities that let an AI agent (Kiro, Claude Code, or any MCP-compatible tool) provision and configure OpenSearch Serverless from natural language.
The agent created a serverless collection, set up IAM roles, configured an Amazon Bedrock connector for Titan text embeddings (1,024-dimension vectors), created ingestion and search pipelines, indexed the movie data, and built a React frontend. All from a single instruction: "Build a search application using the movies data in my local folder."
The search worked as expected. A query for "racing" returned semantically relevant movies, not just exact title matches. Hybrid search handled both concept-based and precise queries. But the infrastructure behavior is what stood out.
After the demo ended and traffic stopped, compute dropped to zero OCUs (OpenSearch Compute Units) within minutes. The collection existed. The data remained indexed. The cost was storage only. When we ran queries again, compute spun up in seconds and the collection responded at full speed.
For teams evaluating hybrid search, the Bedrock integration and automated pipeline setup remove most of the friction. You can prototype a semantic search system in an afternoon and pay nothing for compute between prototype sessions.
Where this matters most
Agentic AI workloads are the clearest fit. Agents are bursty and unpredictable by nature: an agent might fire 50 search queries in 10 seconds during a research task, then go silent for an hour. Provisioning for peak agent traffic wastes money. Provisioning for average risks latency during bursts. Scale-to-zero with seconds-fast autoscaling handles both without you thinking about it.
The same economics apply anywhere traffic is uneven or intermittent. Dev/test environments sit idle 90% of the time, burning compute that nobody is using. QA workloads run a test suite for a few minutes each day, then sit dark until tomorrow. Staging environments spin up for weekly deploys and do nothing in between. Internal tools see a flurry of activity during business hours and silence overnight. Even production search has genuine off-peak hours where query volume drops to a fraction of daytime traffic. In every one of these cases, you are paying for compute capacity that has nothing to do. OpenSearch Serverless makes that idle compute disappear from your bill.
The independent scaling of indexing and search unlocks a pattern that was previously expensive: continuous ingest (keeping your search index fresh from a database CDC stream) without paying for search compute during low-traffic hours. Your data stays current. Your search costs reflect actual query volume, not ingest activity.
Compared to provisioning OpenSearch Service domains for peak capacity, OpenSearch Serverless can deliver up to 60% lower cost for workloads with variable traffic patterns. The savings come from not paying for compute that has nothing to do.
The chmod +x moment
OpenSearch Serverless makes search executable without infrastructure work. You create a collection, point your data pipeline at it, and start querying. Scale happens underneath you. Cost tracks what you actually use. You do not monitor OCU utilization graphs. You do not wake up to adjust capacity for a Monday morning traffic pattern that differs from Sunday. The infrastructure disappears into the background where it belongs, and you spend your time on the search experience your users see rather than the plumbing they never should.
If you are currently sizing clusters, managing OCU minimums from the previous generation, or over-provisioning because you cannot predict traffic, the math just changed. Check your CloudWatch utilization graphs. If average utilization is below 50%, you are paying for compute that OpenSearch Serverless would scale away automatically.
Documentation: Amazon OpenSearch Serverless (https://aws.amazon.com/opensearch-service/features/serverless/)
Top comments (0)