Amazon Bedrock is AWS's managed service for foundation models, and Claude is one of the model families it offers. "Managed service" means AWS operates the infrastructure: you never provision servers, patch anything, or think about capacity hardware. You call an API inside your AWS account, and the request is billed, logged, and access-controlled with the same machinery your teams already use for S3 or RDS. That familiarity is Bedrock's core enterprise appeal — Claude becomes another AWS service rather than a new vendor relationship.
The access model: enable, then invoke
Bedrock does not expose every model to every account by default. Before your first call, someone with the right permissions enables access to the Claude models your organization wants to use, in the regions where you want to use them. This is an administrative step, not a technical one, and it is where week-one delays usually happen — build it into your plan rather than discovering it on demo day.
Once enabled, invocation is an ordinary AWS API call. There are no separate Claude API keys to issue, store, or rotate: requests are authenticated with standard AWS credentials, which for production workloads means IAM roles attached to your compute, not long-lived keys.
Identity and access with IAM
Because Bedrock sits behind AWS IAM, everything your organization already does for access governance applies. You can scope which teams may invoke which models, separate permissions by environment, and require roles rather than user keys for anything automated. Usage shows up in your AWS audit trail like any other service activity, which makes the security review conversation dramatically shorter than onboarding a standalone AI vendor.
The practical patterns — separate policies for dev, staging, and production, scoped to specific models — are covered step by step in Setting Up IAM Permissions for Claude on Bedrock.
Model IDs and the SDK
Bedrock is the one platform where Claude model IDs carry an anthropic. prefix: you request anthropic.claude-sonnet-5 where every other platform uses the bare claude-sonnet-5. The official Python SDK handles Bedrock's authentication and request format through a dedicated client class; install it with pip install -U "anthropic[bedrock]".
from anthropic import AnthropicBedrockMantle
client = AnthropicBedrockMantle(aws_region="us-east-1")
message = client.messages.create(
model="anthropic.claude-sonnet-5",
max_tokens=1024,
messages=[{"role": "user",
"content": "Summarize the attached ticket in two sentences."}],
)
print(message.content)anthropic. prefix is the number-one source of copy-paste confusion. Code samples from Anthropic's docs or from Vertex/Foundry tutorials use bare model IDs and will fail on Bedrock unchanged. See Why Model IDs Differ Across Platforms.Quotas, regions, and day-two operations
Bedrock meters usage with quotas and rate limits per account and region, and the limits you get by default are sized for experimentation, not launch traffic. Treat quota planning as a launch-blocking task: measure your expected requests and tokens per minute during the pilot, then request increases well before go-live. Quotas and Rate Limits covers the process across platforms.
Region choice matters twice: the model must be enabled in the region you call, and your latency and data-location story depends on where those calls land. List pricing on the marketplace matches Anthropic's first-party list prices; if you have meaningful volume, committed-use discounts are negotiated with AWS as part of your existing cloud agreement, not with Anthropic.
What Bedrock does not have
Bedrock covers the core well: the Messages API, streaming, tool use, vision, adaptive thinking, and prompt caching are all available. But as of July 2026 several newer capabilities are absent: the Batch API, the Files API, the code execution tool, the web fetch tool, and the web search tool are not offered on Bedrock. Managed Agents are likewise unavailable (they exist only on the first-party API and Claude Platform on AWS).
Whether this matters depends entirely on your roadmap. A support-summarization or document-extraction workload will never notice. A team planning large overnight batch jobs or agentic features should read The Feature Gaps before committing — and consider Claude Platform on AWS, which stays inside the AWS boundary but tracks the first-party API with same-day parity.
Where to go next
Ready to try it? Your First Claude Call on Amazon Bedrock gets you from console to working response, and the IAM setup guide makes the access story production-grade. For the cross-platform picture, see the platforms comparison.