Overview
Amazon Bedrock is a fully managed service that provides access to foundation models from multiple providers (Anthropic, Amazon, Meta, Mistral and others) through a unified AWS API. It integrates natively with AWS IAM, VPC, CloudWatch, and S3, making it ideal for enterprise workloads requiring compliance, security controls, and AWS-native data pipelines.
Setup
pip install boto3
# Configure credentials (SSO or an IAM role is preferable to long-lived keys)
aws configure sso
export AWS_DEFAULT_REGION=us-east-1
Access to most serverless models is on by default once your IAM role has the AWS Marketplace permissions (aws-marketplace:Subscribe, Unsubscribe, ViewSubscriptions) and the account has a valid payment method. Anthropic models on the bedrock-runtime endpoint also need the one-time first-use form (it does not apply to the bedrock-mantle endpoint). The first call can fail with AccessDeniedException for up to about 15 minutes while the subscription finishes. Some models need extra account-level access from AWS Sales.
Available Models
Model IDs, Regions and lifecycle state change often. List what your account and Region can call, and read the modelLifecycle field before you pin an ID:
import boto3
bedrock = boto3.client("bedrock", region_name="us-east-1") # control plane, not bedrock-runtime
for m in bedrock.list_foundation_models(byProvider="anthropic")["modelSummaries"]:
print(m["modelId"], m["modelLifecycle"]["status"])
| Model ID | Notes |
|---|---|
anthropic.claude-fable-5-1, anthropic.claude-opus-5-5, anthropic.claude-sonnet-5-5, anthropic.claude-haiku-5-5 | Current Claude line. No ARN version suffix. Call through InvokeModel or the Messages endpoint below. |
anthropic.claude-haiku-4-5-20251001-v1:0 | Older Claude. Needs an inference profile (global., us., eu. prefix); the bare ID fails with an HTTP 400 about on-demand throughput. |
meta.llama3-3-70b-instruct-v1:0 | Llama; already Legacy (EOL March 30, 2027). |
Claude 3, 3.5 and 3.7 models, Claude Sonnet 4.5 and Opus 4.1 are Legacy with end-of-life dates between November 2026 and April 2027; Cohere Command R+ and Llama 3.1 405B are already past EOL. A Legacy model stops accepting new customers, and existing ones can lose access after 15 days of inactivity. Check the Bedrock model lifecycle page before using any ID from an older tutorial.
Instructions
Converse API (Recommended)
The Converse API is the unified chat interface for all Bedrock models:
import boto3
bedrock = boto3.client("bedrock-runtime", region_name="us-east-1")
response = bedrock.converse(
modelId="global.anthropic.claude-haiku-4-5-20251001-v1:0",
messages=[{"role": "user", "content": [{"text": "Explain AWS Lambda in simple terms."}]}],
system=[{"text": "You are an AWS solutions architect for a fintech team."}],
inferenceConfig={"maxTokens": 1024, "temperature": 0.7},
)
print(response["output"]["message"]["content"][0]["text"])
print(response["usage"]) # token counts
Streaming with Converse
import boto3
bedrock = boto3.client("bedrock-runtime", region_name="us-east-1")
response = bedrock.converse_stream(
modelId="global.anthropic.claude-haiku-4-5-20251001-v1:0",
messages=[{"role": "user", "content": [{"text": "Write a Python quicksort implementation."}]}],
)
for event in response["stream"]:
if "contentBlockDelta" in event:
delta = event["contentBlockDelta"]["delta"]
if "text" in delta:
print(delta["text"], end="", flush=True)
print()
Claude Sonnet/Opus 5.x through InvokeModel or the Messages endpoint
The current Claude models are documented for InvokeModel on bedrock-runtime and for a native Messages API on the bedrock-mantle endpoint. Use anthropic.claude-sonnet-5-5 as the model ID. Do not assume the Converse examples here accept them; check the model's page in Bedrock first.
pip install -U "anthropic[bedrock]"
from anthropic import AnthropicBedrockMantle
client = AnthropicBedrockMantle(aws_region="us-east-1") # SigV4 via the standard AWS credential chain
message = client.messages.create(
model="anthropic.claude-sonnet-5-5",
max_tokens=1024,
messages=[{"role": "user", "content": "Summarize the CAP theorem in three sentences."}],
)
print(next(b.text for b in message.content if b.type == "text"))
That endpoint has no Files API, server-side tools (web search, code execution), MCP connector or Message Batches.
InvokeModel API (Raw)
Direct access with the model-specific body. Claude's body uses anthropic_version:
import boto3
import json
bedrock = boto3.client("bedrock-runtime", region_name="us-east-1")
body = {
"anthropic_version": "bedrock-2023-05-31",
"max_tokens": 1024,
"messages": [{"role": "user", "content": "What is the capital of France?"}],
}
response = bedrock.invoke_model(
modelId="global.anthropic.claude-haiku-4-5-20251001-v1:0",
body=json.dumps(body),
contentType="application/json",
)
result = json.loads(response["body"].read())
print(result["content"][0]["text"])
Image input
In Converse, an image is a content block with raw bytes (no base64): {"image": {"format": "png", "source": {"bytes": open("diagram.png", "rb").read()}}}, placed next to a {"text": ...} block in the same user message.
Examples
Tool Use (Function Calling)
import boto3
bedrock = boto3.client("bedrock-runtime", region_name="us-east-1")
tools = [
{
"toolSpec": {
"name": "query_database",
"description": "Execute a SQL query against the production database",
"inputSchema": {
"json": {
"type": "object",
"properties": {
"sql": {"type": "string", "description": "SQL query to execute"},
"database": {"type": "string", "description": "Database name"},
},
"required": ["sql"],
}
},
}
}
]
messages = [{"role": "user", "content": [{"text": "How many active users do we have?"}]}]
response = bedrock.converse(
modelId="global.anthropic.claude-haiku-4-5-20251001-v1:0",
messages=messages,
toolConfig={"tools": tools},
)
# Handle tool use
if response["stopReason"] == "tool_use":
tool_use = next(b for b in response["output"]["message"]["content"] if "toolUse" in b)
# tool_use["toolUse"]["name"] / ["input"] say what to run; return the result
messages.append(response["output"]["message"])
messages.append({
"role": "user",
"content": [
{
"toolResult": {
"toolUseId": tool_use["toolUse"]["toolUseId"],
"content": [{"json": {"count": 12483, "active_last_30d": 8921}}],
}
}
],
})
final = bedrock.converse(
modelId="global.anthropic.claude-haiku-4-5-20251001-v1:0",
messages=messages,
toolConfig={"tools": tools},
)
print(final["output"]["message"]["content"][0]["text"])
Knowledge Bases for RAG
import boto3
# Knowledge Base RAG — Bedrock manages embedding and retrieval
account_id = boto3.client("sts").get_caller_identity()["Account"]
bedrock_agent = boto3.client("bedrock-agent-runtime", region_name="us-east-1")
retrieve_response = bedrock_agent.retrieve(
knowledgeBaseId="KB123456789",
retrievalQuery={"text": "What is our refund policy?"},
retrievalConfiguration={
"vectorSearchConfiguration": {"numberOfResults": 5}
},
)
contexts = [r["content"]["text"] for r in retrieve_response["retrievalResults"]]
retrieve_and_generate = bedrock_agent.retrieve_and_generate(
input={"text": "What is our refund policy?"},
retrieveAndGenerateConfiguration={
"type": "KNOWLEDGE_BASE",
"knowledgeBaseConfiguration": {
"knowledgeBaseId": "KB123456789",
"modelArn": f"arn:aws:bedrock:us-east-1:{account_id}:inference-profile/global.anthropic.claude-haiku-4-5-20251001-v1:0",
},
},
)
print(retrieve_and_generate["output"]["text"])
Guardrails for Content Safety
import boto3
bedrock = boto3.client("bedrock-runtime", region_name="us-east-1")
# Apply a guardrail to filter content
response = bedrock.converse(
modelId="global.anthropic.claude-haiku-4-5-20251001-v1:0",
messages=[{"role": "user", "content": [{"text": "My card number is 4111 1111 1111 1111, can you store it?"}]}],
guardrailConfig={
"guardrailIdentifier": "gr7k2m9x4p1q",
"guardrailVersion": "DRAFT", # or "1", "2", etc.
"trace": "enabled",
},
)
# stopReason tells you whether the guardrail stepped in
if response["stopReason"] == "guardrail_intervened":
print("Blocked:", response["output"]["message"]["content"][0]["text"])
print("Why:", response["trace"]["guardrail"]["inputAssessment"])
IAM Policy for Bedrock
{
"Version": "2012-10-17",
"Statement": [
{
"Effect": "Allow",
"Action": [
"bedrock:InvokeModel",
"bedrock:InvokeModelWithResponseStream"
],
"Resource": [
"arn:aws:bedrock:us-east-1:111122223333:inference-profile/global.anthropic.claude-haiku-4-5-20251001-v1:0",
"arn:aws:bedrock:*::foundation-model/anthropic.claude-haiku-4-5-20251001-v1:0"
]
}
]
}
Guidelines
- Use the Converse API for new integrations with models that support it.
Converseneeds thebedrock:InvokeModelpermission andConverseStreamneedsbedrock:InvokeModelWithResponseStream; there are no separatebedrock:Converse*actions. - Newer models are served through inference profiles. Pass the profile ID (
global.,us.,eu.); the policy must allow the profile ARN and the foundation-model ARN in every destination Region, and an SCP that blocks one destination Region fails the request. Regional endpoints cost 10 percent more than global ones for Claude. - Bedrock processes data in the selected AWS region — choose for data residency compliance.
- Knowledge Bases handle chunking, embedding, and retrieval automatically with OpenSearch Serverless.
- Guardrails can block harmful content, PII, and off-topic queries before they reach the model.
- Use
converse_streamfor user-facing features to reduce perceived latency.