Elasticsearch Serverless billing dimensions
Elasticsearch Serverless projects are priced based on consumption of the underlying infrastructure that supports your use case with the performance characteristics you need.
Your bill is made up of project charges, organization charges, and add-ons. This page also explains how to manage costs related to project charges.
Refer to Serverless project billing dimensions and the Elasticsearch Serverless pricing page for more details.
Project charges are based on the capabilities you use in your Elasticsearch Serverless project, starting with VCUs and data storage. Some features are also priced differently depending on the project type, such as Agent Builder and cross-project search.
Measurements are in virtual compute units (VCUs). Each VCU represents a fraction of RAM, CPU, and local disk for caching.
The number of VCUs you need is determined by:
- Volume and ingestion rate of your data
- Number of indices, for ingest VCUs
- Data retention requirements
- Search query volume
- Search Power setting
- Machine learning usage
Elasticsearch Serverless uses the following VCU types:
Ingest: The VCUs used to index incoming documents. Ingest VCUs account for compute resources consumed for ingestion. This is based on ingestion rate and amount of data ingested at any given time. Transforms and ingest pipelines also contribute to ingest VCU consumption.
Because a small amount of memory is allocated for every index, the number of indices also affects ingest VCU consumption. A high index count can raise Ingest VCU use when the ingest rate is unchanged. To reduce this overhead, refer to Indexing strategies.
Ingest VCU billing scales to zero during periods with no ingest activity.
Search: The VCUs used to return search results with the latency and queries per second (QPS) you require. Search VCUs are calculated as a factor of the compute resources needed to run search queries, search throughput, and latency. Search VCUs are not charged per search request. Instead, they are a factor of the compute resources that scale up and down based on amount of searchable data, search load (QPS), and performance (latency and availability).
Unlike ingest, Search VCU billing does not scale to zero during idle periods. A baseline of search capacity is always provisioned so your data remains instantly queryable, and you are billed a reduced rate for this baseline when no search activity is occurring. The baseline is determined primarily by your Search Power setting and your search-ready dataset size.
If you set Minimum Search VCUs, the Search tier does not scale below that number of VCUs. Search VCUs provisioned through this setting are priced at the same per-VCU rate as any other Search VCU consumption.
Machine learning: The VCUs used to perform inference, NLP tasks, and other ML activities. ML VCUs are a factor of the models deployed and number of ML operations such as inference for search and ingest. ML VCUs are typically consumed for generating embeddings during ingestion and during semantic search or reranking.
Models that run on the Elastic Inference Service are billed based on tokens as an add-on, not as ML VCUs.
Elasticsearch Serverless projects store data in the Search AI Lake. You are charged per GB of stored data at rest. Note that if you perform operations at ingest such as vectorization or enrichment, the size of your stored data will differ from the size of the original source data.
Elastic Agent Builder enables you to create AI agents that assist with data exploration, analysis, and automated tasks within your Elasticsearch Serverless project. Agent usage is billed in addition to VCU-based dimensions, based on the number of executions completed in your project over the course of a month.
Each execution represents one conversational turn — a user input and the agent's non-error response. Complex turns that exceed 50,000 input tokens count as additional executions to reflect the heavier processing involved. For example, a turn that uses 120,000 input tokens would count for 3 Agent Executions. Error responses are not billed. When an agent triggers a workflow as part of its response, the workflow execution is metered separately under Workflow Executions.
A free allocation of 1,000 executions per month is included. Volume tier reductions apply at higher usage levels. Each tier's rate applies only to the executions in its range.
Volume tier usage resets at the start of each month. As a result, after your free allocation is used, the same usage can cost more early in the month than at the end of the previous one. Refer to the Elasticsearch Serverless pricing page for specific rates and tier breakpoints.
Cross-project search enables you to search across multiple Serverless projects from a single origin project.
Cross-project search charges are billed to the origin project.
Usage is metered and charged based on the following factors:
- Retained data volume: The volume of data retained in your linked projects each month. The rate depends on the type of your origin project and the type of each linked project, such as Observability, Security, Elasticsearch, or Elasticsearch Vector Database. For Observability linked projects, the rate also depends on the type of retained data.
- CPS queries: Cross-project search queries incur a separate data egress charge for data transferred as a result of those queries. This appears as a separate line item on your bill.
For current rates, refer to the Elastic Cloud Pricing page. You can also create a Serverless estimate.
Elasticsearch Serverless costs follow your resource consumption across search, ingest, ML, and storage. To balance performance with spend for your workload, you can adjust the project settings, data layout, and indexing practices described in this section.
Search Power controls the speed of searches against your data and the size of the baseline resources kept provisioned for your project. With Search Power, you can improve search performance by adding more resources for querying or you can reduce provisioned resources to cut costs.
Minimum Search VCUs lets you pre-scale the Search tier to a baseline level of capacity ahead of anticipated traffic. When you configure this setting, you are billed for at least that number of Search VCUs at all times, including when search load is low or the project is idle. Autoscaling can still provision Search VCUs above the minimum when your Search Power setting, searchable data, or search load require it, and you are also billed for those additional VCUs for as long as they remain provisioned. All Search VCUs are billed at the same per-VCU rate.
After it's been configured, the minimum remains in effect, and continues to be billed, until you lower or disable it in your project settings or through the API. Lowering the minimum might not reduce your Search VCUs immediately, because capacity stays provisioned for as long as search load requires it.
To avoid paying for capacity you don't need, set the minimum no higher than your workload requires, and lower or disable it when you no longer need it.
The search-ready dataset size for a project is 100% of the index size for regular indices and the data within the Boost Window for time series data. This size is an important factor, along with your Search Power setting, that determines the baseline resources provisioned for your project, and therefore the reduced rate billed during idle periods. During active search, Search VCU consumption scales up based on query load; during idle periods, it drops to a reduced rate proportional to the search-ready dataset size.
For regular indices, reducing the search-ready dataset size means reducing what you store and search: consolidating small indices, removing unused data, or moving historical data out of the project. For time series data, shortening the Search Boost Window is typically the most direct lever: data outside the window remains fully searchable from the Search AI Lake but has a minimal impact on the baseline capacity, lowering both active and idle Search VCU costs.
Trained model autoscaling is always enabled and cannot be disabled, ensuring efficient resource usage, reduced costs, and optimal performance without manual configuration.
Trained model deployments automatically scale down to zero allocations after 24 hours without any inference requests. When they scale up again, they remain active for 5 minutes before they can scale down. During these cooldown periods, you will continue to be billed for the active resources.
Consider your indexing strategies and how they might impact overall VCU usage and costs. To ensure optimal performance and cost-effectiveness for your project, it's important to consider how you structure your data. You can control ingest VCU usage by managing the indexing load and how data is distributed across indices:
Bulk indexing: Ingest VCU consumption scales with your indexing load. After 15 minutes with no ingest activity, ingest VCU consumption scales down to zero.
Where your use case allows, batch documents into fewer, larger
_bulkrequests.Index count and size: Consolidate small indices for better efficiency. In general, avoid a design where your project contains hundreds of very small indices, specifically those under 1GB each. Avoiding small indices is important because every index in Elasticsearch has a certain amount of resource overhead. Elasticsearch needs to maintain metadata for each index to keep it running smoothly. When you have a very large number of small indices, the combined overhead from all of them can consume more CPU resources than if the same data were stored in fewer, larger indices. Higher resource consumption can lead to higher costs and potentially impact the overall performance of your project.
If your use case naturally generates many small, separate streams of data, the recommended approach is to implement a process to consolidate them into fewer, larger indices. This practice leads to more efficient resource utilization. By grouping your data into larger indices, you can ensure a more performant and cost-efficient experience with Elasticsearch Serverless.
When you use the API to create projects, be aware that the optimized_for option affects the VCU allocation and costs.
The general_purpose option is suitable for most search use cases. For example, it is the right profile for full-text search, sparse vectors, and dense vectors that use compression such as BBQ. It is used by default when you create projects from the UI.
The vector option is recommended only for uncompressed dense vectors (dense_vector fields with int4 or int8 quantization strategies) and high dimensionality. Though the per VCU cost is the same for general purpose and vector profiles, the latter allocates more VCUs for searchable data. This leads to higher VCU consumption in order to improve the performance for uncompressed vector data.
These charges apply regardless of which project type you use, and include data out and support.
Data out accounts for all of the traffic coming out of a serverless project. This includes search results, as well as monitoring data sent from the project. The same rate applies regardless of the destination of the data, whether to the internet, another region, or a cloud provider account in the same region. Data coming out of the project through AWS PrivateLink, GCP Private Service Connect, or Azure Private Link is also considered data out.
Data transferred as a result of cross-project search queries incurs a separate data egress charge. This appears as a separate line item on your bill.
If your subscription level is Standard, there is no separate charge for Support reflected on your bill. If your subscription level is Gold, Platinum, or Enterprise, a charge is made for Support as a percentage (%) of the Elastic Consumption Units (ECUs). To find out more about our support levels, go to https://www.elastic.co/support.
Add-ons are capabilities that are available across your Serverless projects. They're billed based on usage, in addition to the other billing dimensions of your project.
Workflows enable you to automate multi-step processes within your Elasticsearch project.
In addition to the VCU-based billing dimensions above, workflow usage is billed based on the number of executions completed successfully in your project over the course of a month. Each execution represents one end-to-end run of a workflow. Failed executions are not billed.
A free allocation of 10,000 executions per month is included. Volume tier reductions apply at higher usage levels. Each tier's rate applies only to the executions in its range.
Volume tier usage resets at the start of each month. As a result, after your free allocation is used, the same usage can cost more early in the month than at the end of the previous one. Refer to the Elasticsearch Serverless pricing page for specific rates and tier breakpoints.
The Elastic Inference Service is charged based on the tokens used with its models, including embedding models, rerankers, and Elastic Managed LLMs. For embeddings and rerankers, usage is billed per million input tokens sent to the models. For LLMs, usage is billed per 1 million input tokens, output tokens, and cached tokens, depending on the model.
Elastic Managed LLMs can power AI features in Elasticsearch Serverless projects, such as Agent Builder and AI Assistant, and are enabled by default.