Amazon Bedrock Service Tiers

Amazon Bedrock Service Tiers

The GenAI IDP solution supports Amazon Bedrock service tiers through model ID suffixes, allowing you to optimize for performance and cost by selecting different service tiers for model inference operations.

Overview

Amazon Bedrock offers three service tiers for on-demand inference:

Tier	Performance	Cost	Best For
Priority	Fastest response times	Premium pricing (~25% more)	Customer-facing workflows, real-time interactions
Standard	Consistent performance	Regular pricing	Everyday AI tasks, content generation
Flex	Variable latency	Discounted pricing	Batch processing, evaluations, non-urgent workloads

Model ID Suffix Format

Service tiers are specified using model ID suffixes:

Format: <base-model-id>[:<service-tier>]

Examples:

• us.amazon.nova-2-lite-v1:0 → Standard tier (default, no suffix)
• us.amazon.nova-2-lite-v1:0:flex → Flex tier
• us.amazon.nova-2-lite-v1:0:priority → Priority tier
• global.amazon.nova-2-lite-v1:0:flex → Global model with Flex tier

Available Models with Service Tiers

The following Nova 2 Lite models are available with service tier suffixes:

US Region Models:

• us.amazon.nova-2-lite-v1:0 (Standard - default)
• us.amazon.nova-2-lite-v1:0:flex (Flex tier)
• us.amazon.nova-2-lite-v1:0:priority (Priority tier)

Global Models:

• global.amazon.nova-2-lite-v1:0 (Standard - default)
• global.amazon.nova-2-lite-v1:0:flex (Flex tier)
• global.amazon.nova-2-lite-v1:0:priority (Priority tier)

Configuration

Using Model IDs with Service Tier Suffixes

Simply specify the model ID with the desired tier suffix in your configuration:

classification:
  model: "us.amazon.nova-2-lite-v1:0:priority"  # Fast classification
  # ... other settings

extraction:
  model: "us.amazon.nova-2-lite-v1:0:flex"  # Cost-effective extraction
  # ... other settings

assessment:
  model: "us.amazon.nova-2-lite-v1:0"  # Standard tier (no suffix)
  # ... other settings

How It Works

When you specify a model ID with a service tier suffix:

1. The system parses the model ID to extract the base model and tier
2. The base model ID (without suffix) is passed to the Bedrock API
3. The extracted tier is passed as the serviceTier parameter
4. Example: us.amazon.nova-2-lite-v1:0:flex becomes:
   • Model ID: us.amazon.nova-2-lite-v1:0
   • Service Tier: flex

Web UI Configuration

Selecting Models with Service Tiers

1. Navigate to the Configuration page
2. In each operation section (Classification, Extraction, Assessment, Summarization):
   • Find the “Model” dropdown
   • Select a model with the desired tier suffix:
     • Models ending in :flex use Flex tier
     • Models ending in :priority use Priority tier
     • Models without suffix use Standard tier
3. Changes save automatically

The model dropdown will show all available models including those with tier suffixes.

Use Cases and Recommendations

Priority Tier

When to use:

• Customer-facing document processing
• Real-time classification needs
• Interactive applications
• Time-sensitive workflows

Example configuration:

classification:
  model: "us.amazon.nova-2-lite-v1:0:priority"
  # Fast classification for real-time user uploads

Standard Tier (Default)

When to use:

• General document processing
• Balanced performance and cost
• Most production workloads
• Default choice when unsure

Example configuration:

extraction:
  model: "us.amazon.nova-2-lite-v1:0"
  # No suffix = standard tier

Flex Tier

When to use:

• Batch document processing
• Evaluation and testing workflows
• Non-urgent background jobs
• Cost optimization scenarios

Example configuration:

extraction:
  model: "us.amazon.nova-2-lite-v1:0:flex"
  # Cost-effective for batch processing

Mixed Tier Strategy

You can use different tiers for different operations:

classification:
  model: "us.amazon.nova-2-lite-v1:0:priority"
  # Fast classification for user experience

extraction:
  model: "us.amazon.nova-2-lite-v1:0:flex"
  # Cost-effective extraction (can tolerate latency)

assessment:
  model: "us.amazon.nova-2-lite-v1:0"
  # Standard tier for assessment

Performance Expectations

Based on AWS documentation:

• Priority: Up to 25% better OTPS (Output Tokens Per Second) latency vs Standard
• Standard: Consistent baseline performance
• Flex: Variable latency, suitable for batch workloads where speed is less critical

Actual performance varies by:

• Model size and complexity
• Request payload size
• Current system load
• AWS region

Cost Implications

• Priority: ~25% premium over Standard pricing
• Standard: Baseline pricing (reference point)
• Flex: Discounted pricing compared to Standard

Important: Always refer to the AWS Pricing Calculator for current pricing information.

Monitoring

CloudWatch Logs

Service tier information appears in CloudWatch logs:

Using service tier: flex
Extracted service tier 'flex' from model ID. Using base model ID: us.amazon.nova-2-lite-v1:0

Metrics

CloudWatch metrics include service tier as a dimension, allowing you to:

• Track usage by tier
• Compare performance across tiers
• Analyze cost by tier

Troubleshooting

Model ID Not Recognized

Problem: Model with tier suffix not working

Solution: Verify the model ID format:

• Correct: us.amazon.nova-2-lite-v1:0:flex
• Incorrect: us.amazon.nova-2-lite-v1:0-flex
• Incorrect: us.amazon.nova-2-lite-v1:0_flex

Service Tier Not Applied

Problem: Service tier not being used

Solution: Check CloudWatch logs for:

• “Extracted service tier” messages
• “Using service tier” messages
• Verify model ID includes tier suffix

Invalid Tier Suffix

Problem: Using unsupported tier suffix

Solution: Only flex and priority suffixes are supported. Standard tier is the default (no suffix needed).

Best Practices

1. Start with Standard: Use standard tier (no suffix) as your baseline
2. Optimize Selectively: Add tier suffixes only where needed
3. Monitor Costs: Track spending by tier using CloudWatch metrics
4. Test Performance: Measure actual latency differences for your workload
5. Use Flex for Batch: Leverage Flex tier for non-urgent batch processing
6. Reserve Priority: Use Priority tier only for latency-sensitive operations

Model Compatibility

Not all Bedrock models support all service tiers. Priority and Flex tiers are supported by:

• Amazon Nova models
• OpenAI models
• Qwen models
• DeepSeek models

If a model doesn’t support the specified tier, the Bedrock API will return an error.

Additional Resources

• AWS Bedrock Service Tiers Documentation
• AWS Pricing Calculator
• GenAI IDP Configuration Guide