Put Bedrock inference costs on the right teams with IAM-based allocation

Amazon Bedrock now supports cost allocation by IAM principal for inference requests through the bedrock-mantle endpoint, giving teams a clearer way to connect usage with users and roles.

Customers can tag IAM principals and analyze inference costs in AWS Cost Explorer or Cost and Usage Report (CUR) 2.0. That makes it easier to understand which teams or workloads are driving Bedrock consumption.

As a result, FinOps teams can improve showback and chargeback for AI usage without relying only on broader account-level cost views.

See the real cost of staying on older OpenSearch versions

AWS extended security and operating system patch coverage for legacy Elasticsearch and OpenSearch engine versions through November 7, 2027. The announcement also outlines Extended Support schedules and an Extended Support surcharge for certain versions.

The longer runway gives customers more time to plan and complete migrations. However, the new support charges mean teams will need to include Extended Support costs in migration planning and ongoing budgets.

That combination gives organizations more flexibility while making the financial impact of delaying an upgrade more visible.

Run native vector search in a serverless DynamoDB model

Amazon DynamoDB now offers native vector search generally, with single-digit millisecond latency and more than 99% recall at any scale.

The capability provides semantic retrieval within DynamoDB’s serverless, pay-for-what-you-use model.

For large-scale vector workloads, keeping retrieval in that model can reduce the infrastructure and operational costs associated with running a separate vector search environment.

Replace unhealthy EC2 instances based on application-level checks

Amazon EC2 introduced application status checks that probe HTTP or HTTPS endpoints every 60 seconds.

The checks integrate with EC2 Auto Scaling, allowing unhealthy instances with application failures to be replaced automatically.

Faster detection can reduce the time that unhealthy instances continue running without serving useful work, helping teams avoid unnecessary infrastructure costs and improve service availability.

Run Bedrock agents on customer-managed EC2 capacity

Amazon Bedrock AgentCore runtime instances are now generally available. Agents can run on customer EC2 instances, including GPU, memory-optimized, and compute-optimized types, while AgentCore manages provisioning and the instance lifecycle.

This gives teams more control over the cost and performance profile of sustained or specialized agent workloads.

It also simplifies capacity planning by keeping provisioning and lifecycle management within AgentCore.

Consolidate more collections in OpenSearch Serverless

Amazon OpenSearch Serverless increased the per-collection-group limit from 1,500 to 10,000 collections. The increase applies to the next-generation engine.

The higher limit supports greater consolidation of multi-tenant collections onto shared OpenSearch Compute Units (OCUs).

Better consolidation can improve compute utilization and lower the cost per collection for large multi-tenant search workloads.