Track AI Model Costs

Learn how to track and analyze costs for specific AI models in your Pika deployment using AWS Cost Explorer and inference profiles.

What You'll Accomplish

By the end of this guide, you will:

Understand how Pika enables model-level cost tracking
Find inference profile costs in AWS Cost Explorer
Filter and analyze costs by model, environment, and tags
Create cost allocation reports for budgeting and forecasting

Prerequisites

Pika backend stack deployed to AWS
Access to AWS Cost Explorer in your account
Basic understanding of AWS cost management

How Pika Enables Model Cost Tracking

Pika automatically creates named inference profiles by copying AWS's built-in Bedrock inference profiles. This allows you to:

Track costs for specific models (e.g., Claude 4.5 Sonnet vs Claude 4.5 Haiku)
Differentiate costs between environments (dev vs prod)
Allocate AI costs to projects, teams, or cost centers

What Are Inference Profiles?

Inference profiles are named wrappers around base Bedrock models. They:

Function identically to the base models
Enable granular cost tracking in AWS Cost Explorer
Are automatically created during deployment
Are tagged with your configured stack tags

Step 1: Understand Your Inference Profile Names

Before searching for costs, you need to know your inference profile names.

Naming Pattern

Inference profiles follow this pattern:

{stackName}-{profileName}

Where:

stackName = Your pika.projNameKebabCase + stage (e.g., pika-test, customer-api-prod)
profileName = Model identifier (e.g., claude-sonnet-4-5, claude-haiku-4-5)

Find Your Stack Name

Check your pika-config.ts:

export const pikaConfig: PikaConfig = {
    pika: {
        projNameKebabCase: 'pika',  // This is your base name
        // ...
    }
};

Add your deployment stage:

If deployed with STAGE=test: stack name is pika-test
If deployed with STAGE=prod: stack name is pika-prod

Current Profile Names

Pika creates three inference profiles per deployment:

Model	Profile Name Suffix
Claude 4 Sonnet	`claude-sonnet-4`
Claude 4.5 Haiku	`claude-haiku-4-5`
Claude 4.5 Sonnet	`claude-sonnet-4-5`

Example for STAGE=test:

pika-test-claude-sonnet-4
pika-test-claude-haiku-4-5
pika-test-claude-sonnet-4-5

Step 2: Access AWS Cost Explorer

Navigate to AWS Cost Explorer:
- Sign in to the AWS Management Console
- Go to Billing and Cost Management
- Click Cost Explorer in the left navigation
Select Date Range:
- Choose a date range that includes your deployment
- Recommendation: Start with "Last 30 days" to see recent costs

Step 3: Filter by Inference Profile Name

Option A: Filter by Service (Recommended)

Filter by AWS Bedrock:
- Click Filters → Service
- Search for and select "Amazon Bedrock"
Group by Usage Type:
- Click Group by → Usage Type
- Look for entries containing your inference profile names
- Example: USE2-Inference:pika-test-claude-sonnet-4-5
Identify Your Costs:
- Usage types will include your inference profile names
- Costs are displayed per profile

Option B: Filter by Tags

If you've configured stack tags, use them for more flexible filtering:

Filter by Project Tag:
- Click Filters → Tag
- Select your project tag key (e.g., Project)
- Choose the value (e.g., CustomerPortal)
Add Service Filter:
- Add Service filter → Amazon Bedrock
Group by Component:
- Click Group by → Tag
- Select component tag
- This shows costs per inference profile

Step 4: Analyze Costs by Model

View Costs Per Model

Create a custom view to compare model costs:

Group by Usage Type:
- Set Group by → Usage Type
- Filter to show only Bedrock costs
Filter by Inference Profile Pattern:
- Look for usage types containing your stack name
- Each model will appear as a separate line
Compare Costs:
- Identify which models are most expensive
- Analyze usage patterns over time

Example Analysis

For a project with stack name customer-api-prod:

Usage Type	Monthly Cost	Percentage
`Inference:customer-api-prod-claude-sonnet-4-5`	$450.23	65%
`Inference:customer-api-prod-claude-haiku-4-5`	$180.15	26%
`Inference:customer-api-prod-claude-sonnet-4`	$62.08	9%

Insights:

Claude 4.5 Sonnet accounts for 65% of AI costs
Consider using Claude 4.5 Haiku for simpler queries
Claude 4 Sonnet has minimal usage

Step 5: Compare Costs Across Environments

Use tags to compare dev vs production costs:

Filter by Environment Tag:
- Click Filters → Tag → Environment
- Select multiple values (e.g., dev, prod)
Group by Environment:
- Click Group by → Tag: Environment
Add Model Grouping:
- Enable secondary grouping: Usage Type
Analyze Differences:
- Compare dev vs prod spending
- Identify opportunities to reduce dev costs
- Ensure prod has appropriate budget allocation

Cost Optimization Patterns

Development Environment:

Higher percentage of Haiku usage (lower cost)
Burst patterns (testing phases)
Candidate for auto-scaling down during off-hours

Production Environment:

More Sonnet usage (higher quality responses)
Consistent usage patterns
Requires stable performance and availability

Step 6: Create Cost Allocation Reports

Build custom reports for stakeholders:

Monthly AI Cost Report

Configure Report:
- Date range: Last 3 months
- Service: Amazon Bedrock
- Group by: Tag (Project), then Usage Type
Save Report:
- Click Save as... → New report
- Name: "AI Model Costs - Monthly"
- Set frequency: Monthly
Share:
- Export to CSV for stakeholders
- Create budget alerts based on trends

Model Efficiency Report

Track cost per model over time:

Configure View:
- Date range: Last 90 days
- Chart type: Line chart
- Group by: Usage Type (inference profiles)
Identify Trends:
- Seasonal usage patterns
- Cost increases after feature launches
- Optimization opportunities

Step 7: Set Up Cost Alerts

Create budgets and alerts for AI spending:

Navigate to AWS Budgets:
- Go to Billing → Budgets
- Click Create budget
Configure Budget:
- Budget type: Cost budget
- Scope: Amazon Bedrock service
- Filter by tags (optional): Project, Environment
Set Thresholds:
- Example: $1,000/month for production AI costs
- Alert at 80% and 100% of budget
- Send notifications to team email
Add Forecasting:
- Enable forecasted alerts
- Get warnings before exceeding budget

Common Cost Analysis Patterns

By Project

Track AI costs per project:

Filter: Tag:Project = "CustomerPortal"
Service: Amazon Bedrock
Group by: Tag:component (inference profile)
Date range: Last 30 days

By Team

Allocate costs to teams:

Filter: Tag:Team = "platform"
Service: Amazon Bedrock
Group by: Usage Type
Date range: Month to date

By Environment

Compare environments:

Filter: Service = Amazon Bedrock
Group by: Tag:Environment, then Usage Type
Date range: Last 7 days

Model Comparison

Compare specific models:

Filter: Service = Amazon Bedrock
Filter: Usage Type contains "claude-sonnet"
Group by: Usage Type
Date range: Last 90 days
Chart: Line chart

Cost Optimization Tips

Use Appropriate Models

High-stakes queries: Claude 4.5 Sonnet (higher quality, higher cost)
Simple queries: Claude 4.5 Haiku (good quality, lower cost)
Bulk processing: Claude 4 Sonnet or Haiku (older, less expensive)

Monitor Usage Patterns

Identify peak usage times
Consider batch processing for non-urgent tasks
Review agent configurations for unnecessary model calls

Environment-Specific Strategies

Development:

Default to less expensive models
Use cheaper models for testing
Implement auto-shutdown for after-hours

Production:

Balance quality and cost
Set up cost anomaly detection
Review usage weekly

Tag Consistently

Proper tagging enables better cost analysis:

Always use consistent tag keys
Include Environment, Project, and Team tags
Document your tagging strategy

Troubleshooting

Inference Profiles Not Appearing in Cost Explorer

Problem: Can't find inference profile costs in Cost Explorer

Solutions:

Verify profiles were created:
- Check AWS Bedrock console → Inference profiles
- Look for profiles matching your naming pattern
Wait for cost data:
- Cost data can take 24-48 hours to appear
- Recent deployments may not show costs yet
Check date range:
- Ensure Cost Explorer date range includes deployment date
- Costs only appear after actual model usage

Tags Not Showing in Cost Explorer

Problem: Configured tags aren't available for filtering