Skip to content
AI.info

The Pulse

AWS details per-user cost tracking and limits for Bedrock AI tools

AWS has published a guide to tracing Bedrock inference costs to individual users of AI development and productivity tools. Its approach uses identity data and billing reports, while separate enforcement options carry different delays.

AWS details per-user cost tracking and limits for Bedrock AI tools

Organizations can trace spending on AI coding and productivity tools to individual users when those tools call models through Amazon Bedrock. A guide published by Amazon Web Services on November 13 explains how to combine identity data, billing reports and usage controls to track and limit that spending.

The approach covers tools including Claude Code, Claude Desktop, OpenAI Codex, ChatGPT Work and OpenCode when their inference requests pass through Bedrock. The guide shows how to use these Bedrock capabilities, configuring identity propagation once for every tool that routes inference through Bedrock.

Identities in billing data

Bedrock records the caller’s AWS Identity and Access Management (IAM) identity and publishes it in AWS Cost and Usage Reports, or CUR 2.0. For federated users, that identity includes a session name; optional identity or session tags can add organizational details such as team and cost center.

When users share a role with the same fixed session name, or requests pass through a gateway’s own role, the report aggregates spending under a single principal. With IAM Identity Center, the session name is the user’s sign-in name, giving each user a separate principal.

The billing report aggregates usage by principal and usage type for each export interval: one hour for an hourly export or one day for a daily export. It does not contain the cost of each request. AWS points users seeking per-prompt detail to model invocation logs or request metadata tagging.

Limits need separate enforcement

The identity data provides cost visibility, while enforcement requires additional controls. AWS describes using AWS Budgets actions to set thresholds filtered by IAM principal tag and apply deny policies.

The guide says Budgets data updates up to three times a day and Bedrock charges can take up to 24 hours to post. Heavy users may therefore overspend by up to a day before that method takes effect.

AWS also describes quota tracking with DynamoDB, which takes effect at the next credential refresh, and an LLM gateway proxy that can block calls individually. The gateway approach has higher operational complexity and adds latency to each request, according to the guide.

Configuration and exceptions

The setup in the AWS guide uses IAM Identity Center with an external identity provider, such as Okta, Microsoft Entra ID or Ping Identity. It also calls for CUR 2.0 with IAM principal data, activated cost allocation tags and an Amazon Athena table to query the report.

Users receive sessions through a permission set granting Bedrock access. The permissions depend on the endpoint: Codex and ChatGPT Work can use bedrock-mantle, while Claude Code and Claude Desktop can use bedrock-runtime. Session tags can carry organizational attributes into the billing data and AWS Cost Explorer.

The guide distinguishes those direct Bedrock calls from Claude Platform on AWS and Claude for Enterprise, which route inference through Anthropic-managed infrastructure. For the enterprise product, per-user usage is visible in Anthropic’s admin console and Usage and Cost API; Anthropic enforces the spending limits.

Sources

Companies in this story

Related stories

Explore

More articles