AWS adds IAM-level cost tracking for Bedrock with Athena and CUDOS
Trace every Bedrock inference to its exact IAM principal using CUR 2.0.
AWS has expanded Amazon Bedrock cost attribution to include per-IAM-principal tracking, letting organizations see exactly which user or application triggered each inference request. Part 2 of this guide walks through the full setup: creating a CUR 2.0 Data Export with the "Include caller identity (IAM principal) allocation data" checkbox enabled, which populates the line_item_iam_principal column and IAM principal tags. This configuration also supports hourly time granularity for maximum detail, though AWS warns that enabling it expands CUR file sizes since usage is broken out into multiple rows per principal. First reports can take up to 24 hours to deliver to S3.
The guide then demonstrates how to query this data with Amazon Athena using standard SQL, no infrastructure to manage. For faster deployment, AWS provides a sample repository (github.com/aws-samples/sample-cur-iam-principal-bedrock-tracking) that automates setup through Claude Code or Kiro-CLI. Alternatively, you can deploy the CUDOS dashboard via AWS CloudFormation, which also creates the Athena database. Once set up, you can aggregate spend by team, project, or tenant using cost allocation tags, and visualize results in CUDOS dashboards tailored to your organizational structure. This approach works for both third-party Bedrock tools like Claude Code or Codex and custom applications, enabling precise chargeback and FinOps.
- CUR 2.0 exports with 'caller identity' data populate line_item_iam_principal and iamPrincipal/ tags for per-request tracking.
- Hourly granularity gives maximum detail but expands CUR file sizes by splitting usage into rows per IAM principal.
- Automated setup available via agent.md repo for Claude Code/Kiro-CLI, or manual Athena queries and CloudFormation-deployed CUDOS dashboards.
Why It Matters
Gives AWS teams precise per-user and per-app Bedrock cost visibility, enabling chargeback and FinOps optimization.