Half 1 launched granular value attribution for Amazon Bedrock. This characteristic mechanically traces each inference request again to the IAM principal that made the decision. It confirmed how the brand new line_item_iam_principal column may give you per-user and per-application visibility. With non-compulsory value allocation tags, it’s also possible to mixture spend by staff, undertaking, or tenant utilizing AWS Value Explorer. With this strategy, you’ll be able to usually monitor utilization on the granularity you need for any Bedrock-powered service or utility, whether or not third-party instruments like Claude Code or Codex or your personal customized builds. This submit reveals find out how to visualize and analyze value attribution for Amazon Bedrock by Amazon Athena queries and CUDOS dashboards.
First, this submit reveals you find out how to arrange Value and Utilization Report (CUR) 2.0 by Knowledge Exports instrumented with IAM principal knowledge. Subsequent, this submit reveals you find out how to question CUR knowledge with Amazon Athena for evaluation. It then introduces the brand new capabilities of CUDOS dashboards, together with granular Bedrock value and utilization knowledge. Whereas Athena supplies flexibility for aggregations, integration with completely different enterprise intelligence (BI) instruments, and chargeback processes, CUDOS provides pre-built visuals tailor-made to your group’s particular construction.
Instance question output exhibiting Amazon Bedrock prices by IAM principal and utilization kind
Organising Value and Utilization Experiences (CUR 2.0)
Earlier than you’ll be able to analyze Amazon Bedrock prices, arrange a CUR 2.0 knowledge export and join it to Amazon Athena.
Conditions
You want the next:
An AWS account with billing console entry.
IAM permissions for Value and Utilization Experiences, S3, and Athena.
An S3 bucket for CUR knowledge storage.
Primary familiarity with SQL and the AWS Administration Console.
(Elective) Claude Code or Kiro-CLI for automated setup.
Ensure you have the required IAM permissions.
For Amazon Bedrock value attribution, you have to allow IAM principal knowledge in your CUR 2.0 export in order that the line_item_iam_principal column and related IAM principal tags are populated.
Making a CUR 2.0 knowledge export with IAM principal knowledge
Observe the directions in Creating a normal knowledge export to arrange your CUR 2.0 export. When configuring the export, make sure that the next choices are a part of your ultimate configuration:
Making a CUR 2.0 commonplace knowledge export with caller identification (IAM principal) allocation knowledge enabled
In Extra export content material, choose the next checkboxes:
Embrace caller identification (IAM principal) allocation knowledge — That is the important setting that populates the line_item_iam_principal column and surfaces IAM principal tags (prefixed with iamPrincipal/) in your value knowledge.
In Knowledge desk configurations:
Time granularity: Choose Hourly for optimum element.
In Knowledge export supply choices:
File versioning: Choose Overwrite current report back to keep away from storing duplicate knowledge.
Necessary: Enabling IAM principal knowledge will increase CUR file sizes as a result of utilization that was beforehand a single row is now expanded into a number of rows, one for every IAM principal that contributed to the utilization. For top-volume workloads with many distinct principals, plan your Amazon S3 storage accordingly and think about Amazon S3 Lifecycle insurance policies for older CUR information.
It might take as much as 24 hours for AWS to ship your first CUR 2.0 report back to the S3 bucket.
Connecting CUR 2.0 to Amazon Athena
With Amazon Athena, you’ll be able to question your CUR knowledge utilizing commonplace SQL with no infrastructure to handle. To streamline this setup, an non-compulsory agent.md ability repo is obtainable that you need to use with AI assistants like Claude Code, Kiro-CLI, or Codex. It automates the end-to-end means of connecting your Athena atmosphere to your CUR knowledge.
Clone the repo (git clone https://github.com/aws-samples/sample-cur-iam-principal-bedrock-tracking).
Launch Claude Code (claude) or Kiro CLI (kiro) on this listing.
Immediate: “Learn agent.md and comply with its workflow to arrange Value and Utilization Report monitoring and run the Amazon Bedrock-by-principal question for the present month.”
You might also comply with the handbook setup steps.
When you plan to deploy the CUDOS dashboard as effectively, you’ll be able to deploy it from AWS CloudFormation. This additionally deploys the Athena question database as a part of the method.
As soon as deployed, here’s a pattern take a look at question within the Athena Question Editor that you need to use:
If this returns rows with IAM principal ARNs and Bedrock utilization sorts, your setup is full and prepared for deeper evaluation.
Athena question patterns for Bedrock value monitoring
Together with your CUR 2.0 knowledge accessible in Athena, now you can reply granular value attribution questions utilizing SQL. This part supplies three progressively refined question patterns that cowl the commonest evaluation situations.
Be aware: Within the following queries, exchange your_cur_table_name with the precise title of your CUR Athena desk (for instance, cid_data_export.cur2).
Question 1: Bedrock prices by IAM principal and utilization kind
This question supplies a whole breakdown of Amazon Bedrock spending by caller identification and mannequin utilization. It solutions the query: “Who is asking which fashions, and the way a lot are they spending?”
Instance output:
line_item_iam_principal
line_item_usage_type
total_tokens
total_cost
arn:aws:sts::123456789012:assumed-role/ChatApp/session-1
USW2-anthropic.claude-opus-4-8-mantle-cache-write-tokens-standard
1629.5
$11.2029
arn:aws:sts::123456789012:assumed-role/DocProcessor/batch-7
USW2-Claude4.6Sonnet-output-tokens
68.579
$1.131
arn:aws:sts::123456789012:assumed-role/ClaudeCode/chat
USW2-Claude4.6Sonnet-cache-write-input-token-count
831.74
$3.4309
arn:aws:iam::123456789012:consumer/alice
USW2-Claude4.6Sonnet-input-tokens
17.33
$0.0572
Evaluation ideas:
Use LIKE patterns corresponding to line_item_usage_type LIKE ‘%Sonnetpercentoutput%’ or %nova% to filter for particular fashions.
The line_item_iam_principal column comprises the total ARN. For assumed roles, the session title after the final / identifies the particular consumer or session.
Question 2: Prices damaged down by identified IAM principal tags
Whenever you’ve tagged your IAM principals with dimensions like staff, undertaking, or costcenter (and activated these tags as value allocation tags), they seem in your CUR 2.0 knowledge as a part of the tags columns, with the iamPrincipal/ prefix adopted by your tag key. This question teams prices by these tags to reply questions like: “How a lot did the engineering staff spend on Bedrock this month?” or “What’s the full Bedrock value for the chatbot undertaking?”
By undertaking:
Instance output:
undertaking
line_item_usage_type
total_tokens
total_cost
data-science
USW2-Claude4.5Sonnet-cache-write-input-token-count
433.893
1.789808625
data-science
USW2-Claude4.6Sonnet-cache-read-input-token-count
5372.659
1.77297747
engineering
USW2-Claude4.5Sonnet-input-tokens
29.481
0.0972873
engineering
USW2-Claude4.5Sonnet-output-tokens
31.102
0.513183
Be aware: This question returns outcomes provided that your IAM principals have been tagged with the related keys and people tags have been activated as value allocation tags.
Question 3: Dynamic tag discovery utilizing UNNEST for unknown tag schemas
In massive organizations, it’s possible you’ll not know prematurely which tags have been utilized throughout all IAM principals. Totally different groups may use completely different tag keys, or new tags could be launched over time. The next instance reveals find out how to dynamically discover utilizing Athena’s UNNEST operate.
This question discovers all IAM principal tags getting used throughout your Bedrock workloads and reveals value allocation for every tag key-value pair:
Actual-world use case: Multi-service value comparability
Take into account a platform staff working a number of AI-powered companies. For instance, the staff may run a doc summarization pipeline (DocProcessor) and a customer-facing chatbot (ChatApp). The staff can assign every service to its personal IAM function. With the question patterns from the previous part, you’ll be able to isolate how the spend for every service traits with the next question:
Instance output:
line_item_iam_principal
line_item_usage_type
total_usage
total_cost
arn:aws:sts::123456789012:assumed-role/ChatApp/session-1
USE1-Claude4.6Sonnet-output-tokens
4,800,000
$72.00
arn:aws:sts::123456789012:assumed-role/ChatApp/session-1
USE1-Claude4.6Sonnet-input-tokens
2,900,000
$8.70
arn:aws:sts::123456789012:assumed-role/DocProcessor/batch-7
USE1-NovaLite-output-tokens
6,100,000
$1.46
arn:aws:sts::123456789012:assumed-role/DocProcessor/batch-7
USE1-NovaLite-input-tokens
3,200,000
$0.19
From this output, the platform staff can reply questions like:
Which utility is the highest contributor to this month’s Bedrock spend? On this instance, ChatApp accounts for over $80 utilizing Claude 4.6 Sonnet, whereas DocProcessor prices below $5 utilizing Nova Lite.
May we cut back prices by utilizing a distinct mannequin for every workload? DocProcessor is already on Nova Lite (acceptable for simple summarization duties), however the staff may consider whether or not ChatApp might deal with some interactions with a lighter mannequin to cut back the $72 output-token value.
Value of Athena queries
You pay just for the queries that you simply run. You might be charged based mostly on the quantity of knowledge scanned by every question. The console shows this info after every question, and this info can also be accessible within the Latest Queries tab within the Athena console.
Athena queries are billed at $5 per TB scanned (with a ten MB minimal per question). As a result of our desk mechanically makes use of hive partition projection on billing_period, queries scoped to a single month solely scan the parquet information in that month’s folder. Scans are usually effectively below 10 MB, which prices about $0.00005 per question (the ten MB minimal).
To maintain prices low, at all times embody a WHERE billing_period = … filter and choose solely the columns you want reasonably than SELECT *.
The Cloud Intelligence Dashboards framework
CUDOS dashboard is a part of the open supply Cloud Intelligence Dashboards (CID) framework, which you’ll deploy in your AWS account utilizing the supplied infrastructure as code (IaC) templates. The framework helps you drive monetary accountability and enhance operational effectivity throughout your AWS organizations. The CUDOS dashboard supplies detailed and actionable insights, supporting data-driven selections for value effectivity throughout your AWS infrastructure.
Amazon Bedrock value and utilization insights in CUDOS
CUDOS model 5.8 introduces a complete Amazon Bedrock part within the AI/ML tab, with full IAM principal value attribution help. The dashboard supplies:
Versatile grouping dimensions: Group your Amazon Bedrock spend by IAM Principal, IAM Principal Tags (corresponding to Venture or Workforce), Mannequin/Useful resource Group, Area, or another value taxonomy fields configured throughout dashboard deployment.
Value-per-million-tokens monitoring: A development line overlaid in your spend chart exhibiting how value per million tokens evolves over time, serving to you measure the affect of mannequin choice modifications or immediate optimization efforts, like caching for instance.
The next determine reveals the Amazon Bedrock Abstract part within the AI/ML tab of the CUDOS dashboard, grouped by IAM principal, with per-principal value breakdown and cost-per-million-tokens development.
CUDOS dashboard exhibiting Amazon Bedrock spend grouped by IAM principal
Interactive drill-down filtering: Select any worth within the top-level spend chart (corresponding to a selected undertaking, principal, or account) and each different visible mechanically filters to that choice, letting you drill from a high-level overview into per-model and per-usage-type element with out navigating away from the dashboard.
Granular mannequin and utilization breakdown: Extra visuals, filtered by the top-level chart, present spend per mannequin, per utilization kind, and cost-per-million-tokens by mannequin, so you’ll be able to determine which fashions and token sorts are driving prices for a given staff or undertaking.
Whenever you change the grouping to IAM Principal Tag Venture and select a selected undertaking (on this instance, “chatbot-v2”), all different visuals filter to point out solely that undertaking’s spend. The visuals break down that spend by mannequin, utilization kind, and unit value traits.
CUDOS dashboard exhibiting Amazon Bedrock spend grouped by IAM principal tag Venture, filtered to the chatbot-v2 undertaking
With these visuals, you’ll be able to shortly reply questions like “Which undertaking is driving probably the most output token prices?”, “Is our chatbot staff utilizing cost-efficient fashions?”, or “How has our value per million tokens modified since we switched from Opus to Sonnet?”, with out writing any SQL.
Getting began with CUDOS
To get began with CUDOS, you’ll be able to discover the Bedrock part in an interactive demo dashboard. Observe the deployment information to arrange CUDOS in your group.
When you’re already utilizing CUDOS, comply with the replace steering to improve to model 5.8. It’s also possible to use the add organizational taxonomy so as to add IAM Principal knowledge to your current CUDOS dashboard.
Clear up
First, drop the Athena desk and AWS Glue database (these are metadata, so no compute is working):
Warning: Dropping the Athena desk and Glue database will take away your means to question CUR knowledge. You will have to recreate these sources following the previous part if you wish to analyze billing knowledge sooner or later.
Second, for those who not want the associated fee knowledge itself, disable the export within the AWS Billing and Value Administration console below Knowledge Exports and empty the S3 prefix it writes to. Be aware that is your uncooked billing historical past, so solely delete it for those who’re positive.
Lastly, filter any Athena question outcomes which have collected:
There aren’t any crawlers, AWS Lambda features, or schedules to delete. Partition projection means the one ongoing value is S3 storage for the CUR information themselves, which is often pennies per 30 days.
For cleansing up CUDOS, seek advice from the CUDOS dashboard teardown directions.
Conclusion and subsequent steps
This two-part sequence walked by a whole toolkit for understanding and managing Amazon Bedrock inference prices:
Half 1 launched granular value attribution: how Amazon Bedrock mechanically captures the IAM principal behind each inference name, and the way you need to use value allocation tags to mixture spend by staff, undertaking, or tenant.
Half 2 (this submit) confirmed you find out how to put that knowledge to work: establishing CUR 2.0 with IAM principal knowledge, querying value patterns in Amazon Athena, and evaluating spend throughout tasks and principals to tell value allocation selections. It additionally introduces the CUDOS dashboard, which supplies a complete AI/ML sheet with related insights for Bedrock.
Activate CUR 2.0 with caller identification knowledge within the AWS Billing console, join it to Athena utilizing the supplied agent.md file, and run your first cost-by-principal question. Monitor Bedrock adoption throughout your group with the CUDOS dashboard.
Concerning the authors
Abhi Shivaditya
Abhi is a Principal Options Architect at AWS, working with strategic international enterprise organizations to facilitate the adoption of AWS companies in areas corresponding to Synthetic Intelligence, distributed computing, networking, and storage. Abhi assists prospects in deploying high-performance machine studying fashions effectively inside the AWS ecosystem.
Brenno Passanha
Brenno is a Senior Technical Account Supervisor. He’s a part of the Cloud Operations Technical Subject Group, specializing in Cloud Monetary Administration. Outdoors of labor, Brenno enjoys elevating his youngsters, touring the world, and creating reminiscences by new experiences.
Yash Yamsanwar
Yash is a Machine Studying Architect at Amazon Net Companies (AWS), the place he designs high-performance, scalable infrastructure for large-scale LLM inference and agentic AI methods. His work spans the total lifecycle of machine studying fashions — from coaching to manufacturing deployment — with a concentrate on optimizing generative AI methods at scale. Yash collaborates carefully with ML analysis groups to push the boundaries of what’s doable with massive language fashions and different frontier machine studying applied sciences.
