Answer First
Amazon Bedrock now supports IAM-principal cost allocation for inference through the bedrock-mantle endpoint. The August 11, 2026 launch extends principal attribution already available for bedrock-runtime.
To make the data operational:
- decide whether the cost owner is a user, role, team, project, application, or cost center;
- preserve that identity through direct calls, federation, or gateways;
- apply and activate appropriate IAM principal or session tags;
- create a CUR 2.0 export with caller identity enabled;
- validate the resulting Cost Explorer and CUR rows after the expected delay;
- reconcile attributed cost to the Bedrock invoice before chargeback.
Attribution is reporting evidence. It does not enforce permission or prevent spend.
What AWS Records
AWS documentation describes these billing fields:
- caller identity in the CUR 2.0
line_item_iam_principalcolumn; - IAM principal tags with an
iamPrincipal/prefix; - IAM user, IAM role, or assumed-role session ARNs;
- grouping or filtering by activated principal tags in Cost Explorer;
- line-item analysis in CUR 2.0.
Useful tag dimensions include team, department, cost center, project, application, and environment. Avoid high-cardinality timestamps, random identifiers, or per-request values as billing tags.
Endpoint and Attribution Choices
| Method | Identity job | Current endpoint fit |
|---|---|---|
| IAM principal attribution | User, role, team, or other identity dimension | bedrock-runtime and bedrock-mantle on documented supported APIs |
| Bedrock Projects | Application or workload | bedrock-mantle |
| Application inference profiles | Application or workload | bedrock-runtime |
| Request metadata | Per-request application context | Check the current API and logging path |
AWS recommends Projects for application-level attribution and IAM principals for user-level attribution within the same account. Do not force one tag to answer both questions.
Setup Sequence
1. Map the caller chain
Document the human or workload identity, identity provider, STS role assumption, gateway, AWS account, endpoint, model, and billing owner. A direct caller and a shared LLM gateway produce different attribution unless the gateway preserves a usable dimension.
2. Apply principal or session tags
Principal tags are static on an IAM user or role. Session tags can vary when a role is assumed and are useful for federated users or shared gateways. AWS says a session tag overrides a principal tag with the same key for that session.
Session tags are fixed at STS AssumeRole time and apply across calls made with those credentials. They appear as aggregated billing data; they are not written to Bedrock model invocation logs, which record the caller identity ARN. If request-level identity is required, assess distinct role sessions or request metadata.
3. Activate billing tags
An IAM principal tag appears for activation only after the tagged principal makes a Bedrock API call. AWS says tag keys can take up to 24 hours to appear and then up to 24 hours to activate. Activation is not retroactive.
4. Create the right CUR 2.0 export
Create a standard CUR 2.0 data export and select Include caller identity (IAM principal) allocation data. If an export existed before this was enabled, AWS says to create a new export; older exports do not gain identity data retroactively.
Enabling principal detail can multiply CUR rows by the number of calling identities for each model, increasing file size, query cost, and storage requirements.
5. Validate with controlled requests
Send small requests through each expected identity path. After billing data arrives, verify:
- endpoint and supported API;
- principal ARN;
- principal or session tag values;
- account, Region, model, usage, and cost;
- Cost Explorer grouping;
- CUR 2.0 row totals against the invoice;
- untagged and shared-role remainder.
AWS notes that cost data can take up to 24 hours to appear after a request, so this is not a real-time enforcement signal.
Shared Gateways: The Main Failure Mode
If every user calls through one gateway execution role, the role—not the upstream person—becomes the default billing identity. AWS documents session tags for passing user or tenant dimensions through role assumption, but re-assuming a role for every request can hit STS quotas. Cache credentials for the intended session lifetime and design cardinality deliberately.
Do not put unnecessary personal data in cost tags. Finance needs a stable allocation dimension, not a copy of every request identity. Use access-controlled mapping when a pseudonymous or organizational key is sufficient.
Cost Governance Beyond Attribution
Add:
- least-privilege IAM and model access;
- per-project budgets and alerts;
- quota and concurrency controls;
- anomaly detection for spend and token use;
- owner and escalation rules for untagged cost;
- tag-change and offboarding processes;
- monthly reconciliation and dispute handling.
Use the Bedrock Mantle cost-allocation readiness tool before declaring chargeback ready.
Frequently Asked Questions
Does Bedrock Mantle support IAM principal cost allocation?
Yes. AWS announced on August 11, 2026 that model inference through the bedrock-mantle endpoint can be attributed by IAM user or role, extending a capability already available for bedrock-runtime.
Where can I analyze Bedrock IAM principal costs?
Activated IAM principal tags can be filtered or grouped in AWS Cost Explorer. CUR 2.0 can include caller identity in line_item_iam_principal and principal tags with the iamPrincipal/ prefix when the export is configured for caller identity data.
Are IAM principal cost tags retroactive?
No. AWS says cost allocation tags apply only to costs incurred after activation. Tags can take time to appear and activate, and an existing CUR 2.0 export must be replaced with a new caller-identity-enabled export for this data.
How do shared gateway roles affect attribution?
A gateway’s execution role is the captured identity unless the design preserves user or tenant context. AWS documents session tags, per-user role sessions, request metadata, and Projects as different mechanisms; select them according to identity and workload needs.
Does cost attribution enforce a budget or access policy?
No. Attribution explains recorded spend after requests occur. IAM authorization, quotas, application controls, budgets, alerts, anomaly detection, and response ownership remain separate controls.