AWS Bedrock

Amazon Bedrock provides access to foundation models (FMs) from AI companies through a unified API. Build generative AI applications with text generation, embeddings, and image generation capabilities.

Table of Contents

• Core Concepts
• Common Patterns
• CLI Reference
• Best Practices
• Troubleshooting
• References

Core Concepts

Foundation Models

Pre-trained models available through Bedrock:

• Claude (Anthropic): Text generation, analysis, coding
• Nova / Titan (Amazon): Text, multimodal, embeddings
• GPT / gpt-oss (OpenAI): Text generation, reasoning
• Llama (Meta): Open-weight text generation
• Mistral: Efficient text generation
• Stable Image (Stability AI): Image generation and editing

Model Access

In commercial Regions, access to all serverless models is enabled by default (no console opt-in). In GovCloud (US), models are still enabled manually on the Model access page (third-party models also in the linked commercial account):

• First invocation of a third-party model auto-subscribes via AWS Marketplace (up to 15 min); caller needs aws-marketplace:Subscribe, Unsubscribe, ViewSubscriptions
• Anthropic models on bedrock-runtime need a one-time use case form per account/org (put-use-case-for-model-access)
• Invoking implies EULA acceptance; to block a model, deny both bedrock:InvokeModel and bedrock:InvokeModelWithResponseStream on it (SCP/IAM); streaming APIs such as ConverseStream use the latter. Denying aws-marketplace:Subscribe alone does not block first use

Endpoints

• Same per-token price on both; auth via SigV4 or Bedrock API key (AWS_BEARER_TOKEN_BEDROCK)
• IAM: bedrock:InvokeModel (runtime) vs bedrock-mantle:CreateInference (mantle)
• Responses API on bedrock-runtime is synchronous only and has no server-side tools

Inference Profiles and Model Lifecycle

• Newer models (e.g. Claude Sonnet 5) have no in-Region on-demand ID on bedrock-runtime: use a geo (us., eu., au.) or global. inference profile ID as modelId
• Lifecycle is Active -> Legacy -> EOL (see modelLifecycle in get-foundation-model). Legacy: no new Provisioned Throughput, fine-tuning, or quota increases; EOL: requests fail
• Model cards list an "EOL no sooner than" date; check before pinning a model ID

Knowledge Bases and Agents

• Managed knowledge bases (type: MANAGED): Bedrock runs storage, indexing, and retrieval. Only type that supports AgenticRetrieveStream (query decomposition, iterative retrieval, optional AgentCore Memory via memoryConfiguration)
• Native multimodal managed KBs embed video/audio/image directly with TwelveLabs Marengo Embed 3.0 (twelvelabs.marengo-embed-3-0-v1:0); query with text via Retrieve only (no RetrieveAndGenerate)
• Bedrock Agents Classic is in maintenance mode: closed to new accounts since July 30, 2026 (CreateAgent/InvokeInlineAgent return 403 without prior 12-month usage), model catalog frozen. Build new agents on Amazon Bedrock AgentCore

Inference Types

Common Patterns

Invoke Model (Text Generation)

AWS CLI:

# Invoke Claude
aws bedrock-runtime invoke-model \
  --model-id us.anthropic.claude-sonnet-5 \
  --content-type application/json \
  --accept application/json \
  --cli-binary-format raw-in-base64-out \
  --body '{
    "anthropic_version": "bedrock-2023-05-31",
    "max_tokens": 4096,
    "messages": [
      {"role": "user", "content": "Explain AWS Lambda in 3 sentences."}
    ]
  }' \
  response.json

# Claude Sonnet 5/Opus 5 think by default: content may start with a thinking block
cat response.json | jq -r '.content[] | select(.type=="text") | .text'

boto3:

import boto3
import json

bedrock = boto3.client('bedrock-runtime')

def invoke_claude(prompt, max_tokens=4096):
    response = bedrock.invoke_model(
        modelId='us.anthropic.claude-sonnet-5',
        contentType='application/json',
        accept='application/json',
        body=json.dumps({
            'anthropic_version': 'bedrock-2023-05-31',
            'max_tokens': max_tokens,
            'messages': [
                {'role': 'user', 'content': prompt}
            ]
        })
    )

    result = json.loads(response['body'].read())
    # Skip thinking blocks (adaptive thinking is on by default for Sonnet 5).
    # max_tokens caps thinking + text, so a truncated response may have no text block.
    if result['stop_reason'] == 'max_tokens':
        print('Truncated at max_tokens: raise it or lower output_config.effort')
    return next((b['text'] for b in result['content'] if b['type'] == 'text'), '')

# Usage
response = invoke_claude('What is Amazon S3?')
print(response)

Streaming Response

import boto3
import json

bedrock = boto3.client('bedrock-runtime')

def stream_claude(prompt):
    response = bedrock.invoke_model_with_response_stream(
        modelId='us.anthropic.claude-sonnet-5',
        contentType='application/json',
        accept='application/json',
        body=json.dumps({
            'anthropic_version': 'bedrock-2023-05-31',
            'max_tokens': 4096,
            'messages': [
                {'role': 'user', 'content': prompt}
            ]
        })
    )

    for event in response['body']:
        chunk = json.loads(event['chunk']['bytes'])
        if chunk['type'] == 'content_block_delta':
            yield chunk['delta'].get('text', '')

# Usage
for text in stream_claude('Write a haiku about cloud computing.'):
    print(text, end='', flush=True)

Generate Embeddings

import boto3
import json

bedrock = boto3.client('bedrock-runtime')

def get_embedding(text):
    response = bedrock.invoke_model(
        modelId='amazon.titan-embed-text-v2:0',
        contentType='application/json',
        accept='application/json',
        body=json.dumps({
            'inputText': text,
            'dimensions': 1024,
            'normalize': True
        })
    )

    result = json.loads(response['body'].read())
    return result['embedding']

# Usage
embedding = get_embedding('AWS Lambda is a serverless compute service.')
print(f'Embedding dimension: {len(embedding)}')

Conversation with History

import boto3
import json

bedrock = boto3.client('bedrock-runtime')

class Conversation:
    def __init__(self, system_prompt=None):
        self.messages = []
        self.system = system_prompt

    def chat(self, user_message):
        self.messages.append({
            'role': 'user',
            'content': user_message
        })

        body = {
            'anthropic_version': 'bedrock-2023-05-31',
            'max_tokens': 4096,
            'messages': self.messages
        }

        if self.system:
            body['system'] = self.system

        response = bedrock.invoke_model(
            modelId='us.anthropic.claude-sonnet-5',
            contentType='application/json',
            accept='application/json',
            body=json.dumps(body)
        )

        result = json.loads(response['body'].read())
        if result['stop_reason'] == 'max_tokens':
            # max_tokens caps thinking + text; don't store a truncated/empty turn
            self.messages.pop()
            raise RuntimeError('Truncated at max_tokens: raise it or lower output_config.effort')
        assistant_message = next(
            (b['text'] for b in result['content'] if b['type'] == 'text'), ''
        )

        self.messages.append({
            'role': 'assistant',
            'content': assistant_message
        })

        return assistant_message

# Usage
conv = Conversation(system_prompt='You are an AWS solutions architect.')
print(conv.chat('What database should I use for a chat application?'))
print(conv.chat('What about for time-series data?'))

List Available Models

# List all foundation models
aws bedrock list-foundation-models \
  --query 'modelSummaries[*].[modelId,modelName,providerName]' \
  --output table

# Filter by provider
aws bedrock list-foundation-models \
  --by-provider anthropic \
  --query 'modelSummaries[*].modelId'

# Get model details (includes modelLifecycle.status)
aws bedrock get-foundation-model \
  --model-identifier anthropic.claude-sonnet-5

Check Model Access

# agreementAvailability.status AVAILABLE / NOT_AVAILABLE, authorizationStatus
aws bedrock get-foundation-model-availability \
  --model-id anthropic.claude-sonnet-5

# Anthropic one-time use case form (base64-encoded JSON:
# companyName, companyWebsite, intendedUsers, industryOption, otherIndustryOption, useCases)
aws bedrock put-use-case-for-model-access --form-data <base64-json>

# Programmatic agreement for third-party models
aws bedrock list-foundation-model-agreement-offers --model-id <model-id>
aws bedrock create-foundation-model-agreement --model-id <model-id> --offer-token <token>

Count Tokens

# Free; returns inputTokens. Not supported for every model (e.g. CRIS-only Claude models)
aws bedrock-runtime count-tokens \
  --model-id anthropic.claude-3-5-haiku-20241022-v1:0 \
  --input '{"converse": {"messages": [{"role": "user", "content": [{"text": "Hello"}]}]}}'

CLI Reference

Bedrock (Control Plane)

Bedrock Runtime (Data Plane)

InvokeModelWithResponseStream and ConverseStream are SDK-only (not in AWS CLI v2).

Bedrock Agent Runtime

InvokeAgent, RetrieveAndGenerateStream, and AgenticRetrieveStream are SDK-only (event streams).

Best Practices

Cost Optimization

• Use appropriate models: Smaller models for simple tasks
• Set max_tokens: Limit output length when possible
• Cache responses: For repeated identical queries
• Batch when possible: Use batch inference for bulk processing
• Monitor usage: Set up CloudWatch alarms for cost
• Global inference profiles: ~10% cheaper than geo profiles when data residency allows
• Thinking tokens bill as output: Claude Sonnet 5/Opus 5 think by default; pass "thinking": {"type": "disabled"} or lower output_config.effort if not needed, and revisit max_tokens (it caps thinking + text)
• Converse batch format: --model-invocation-type Converse keeps one request shape across models
• Cost attribution: Tag IAM principals as cost allocation tags (works on both endpoints)

Performance

• Use streaming: For better user experience with long outputs
• Connection pooling: Reuse boto3 clients
• Regional deployment: Use closest region to reduce latency
• Provisioned throughput: For consistent high-volume workloads
• Endpoint choice: Default to bedrock-runtime; use bedrock-mantle only for mantle-only features

Security

• Least privilege IAM: Only grant needed model access
• VPC endpoints: Keep traffic private
• Guardrails: Implement content filtering
• Audit with CloudTrail: Track model invocations
• Web Search (mantle): Set external_web_access: false to keep Fetch inside the AWS boundary; AmazonBedrockFullAccess lacks bedrock-websearch:ExternalWebAccess, so the default true silently fails Fetch
• Pin active models: Check modelLifecycle and migrate off Legacy models before EOL

IAM Permissions

{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Effect": "Allow",
      "Action": [
        "bedrock:InvokeModel",
        "bedrock:InvokeModelWithResponseStream"
      ],
      "Resource": [
        "arn:aws:bedrock:us-east-1:123456789012:inference-profile/us.anthropic.claude-sonnet-5",
        "arn:aws:bedrock:us-east-1::foundation-model/amazon.titan-embed-text-v2:0"
      ]
    },
    {
      "Effect": "Allow",
      "Action": [
        "bedrock:InvokeModel",
        "bedrock:InvokeModelWithResponseStream"
      ],
      "Resource": "arn:aws:bedrock:*::foundation-model/anthropic.claude-sonnet-5",
      "Condition": {
        "StringEquals": {
          "bedrock:InferenceProfileArn": "arn:aws:bedrock:us-east-1:123456789012:inference-profile/us.anthropic.claude-sonnet-5"
        }
      }
    }
  ]
}

Inference profiles need access to the profile ARN plus the foundation model in every destination Region (list them with aws bedrock get-inference-profile --inference-profile-identifier <id>, models field). SCPs that deny Regions must allow those destinations (or exempt via bedrock:InferenceProfileArn).

Troubleshooting

AccessDeniedException

Causes:

• Missing aws-marketplace:Subscribe on first use of a third-party model (auto-subscription fails; may take ~2 min after fixing)
• Anthropic use case form not submitted
• IAM policy missing bedrock:InvokeModel, or missing destination-Region foundation-model ARNs for an inference profile
• Wrong model ID or region
• Bedrock Agents Classic: "Bedrock Agents is in Maintenance Mode" 403 on CreateAgent/InvokeInlineAgent in accounts without prior usage (use AgentCore)

Debug:

# Check model access status
aws bedrock get-foundation-model-availability \
  --model-id anthropic.claude-sonnet-5

# Test IAM permissions
aws iam simulate-principal-policy \
  --policy-source-arn arn:aws:iam::123456789012:role/my-role \
  --action-names bedrock:InvokeModel \
  --resource-arns "arn:aws:bedrock:us-east-1:123456789012:inference-profile/us.anthropic.claude-sonnet-5"

# The profile ARN can pass while cross-Region routing is still denied: also simulate each
# destination-Region model ARN (models field of get-inference-profile) with the profile context
aws iam simulate-principal-policy \
  --policy-source-arn arn:aws:iam::123456789012:role/my-role \
  --action-names bedrock:InvokeModel \
  --resource-arns "arn:aws:bedrock:us-east-1::foundation-model/anthropic.claude-sonnet-5" \
                  "arn:aws:bedrock:us-west-2::foundation-model/anthropic.claude-sonnet-5" \
  --context-entries '[{"ContextKeyName":"bedrock:InferenceProfileArn","ContextKeyType":"string","ContextKeyValues":["arn:aws:bedrock:us-east-1:123456789012:inference-profile/us.anthropic.claude-sonnet-5"]}]'

ModelNotReadyException

Cause: Model is still being provisioned or temporarily unavailable.

Solution: Implement retry with exponential backoff:

import time
from botocore.exceptions import ClientError

def invoke_with_retry(bedrock, body, max_retries=3):
    for attempt in range(max_retries):
        try:
            return bedrock.invoke_model(
                modelId='us.anthropic.claude-sonnet-5',
                body=json.dumps(body)
            )
        except ClientError as e:
            if e.response['Error']['Code'] == 'ModelNotReadyException':
                time.sleep(2 ** attempt)
            else:
                raise
    raise Exception('Max retries exceeded')

ThrottlingException

Causes:

• Exceeded per-model tokens-per-minute (input + output combined on bedrock-runtime) or tokens-per-day quota
• RPM quota (model-specific; some models have none)
• Too many concurrent requests

Solutions:

• Request quota increase (request "Cross-Region InvokeModel tokens per minute for <model>" to cover TPM/TPD together; not granted for Legacy models)
• Lower max_tokens: it affects quota deduction
• Use a cross-Region inference profile for higher throughput
• Implement exponential backoff
• Consider provisioned throughput

ValidationException

Common issues:

• Invalid model ID, or model is EOL
• Error mentions on-demand throughput not supported for the model ID: use an inference profile ID (us./global. prefix)
• Malformed request body
• max_tokens exceeds model limit
• thinking.type: "enabled" with budget_tokens on models that only accept adaptive/disabled (e.g. Claude Sonnet 5)
• output_config.format (structured outputs) sent to bedrock-mantle (use Converse/InvokeModel on bedrock-runtime)

Debug:

# Check model-specific requirements
aws bedrock get-foundation-model \
  --model-identifier anthropic.claude-sonnet-5 \
  --query 'modelDetails.[inferenceTypesSupported,modelLifecycle.status]'

References

• Bedrock User Guide
• Bedrock API Reference
• Bedrock Runtime API
• Model Parameters
• Bedrock Pricing
• Endpoints
• Models at a Glance
• Model Lifecycle
• Agents Classic Maintenance Mode