Amazon Bedrock gives you access to Claude models through AWS infrastructure. Instead of calling the Anthropic API directly, you send requests through Bedrock, which means your data stays within your AWS account and benefits from the security, compliance, and networking features you already have in place.
This article covers how to set up Claude on Bedrock, how pricing compares to the direct API, and when Bedrock is the right choice.
What Is Amazon Bedrock?
Amazon Bedrock is a fully managed service that provides access to foundation models from multiple providers — Anthropic (Claude), Meta (Llama), Amazon (Titan), and others — through a unified API. You do not need to provision servers or manage infrastructure. You call an API endpoint, and Bedrock handles the inference.
The key difference from calling Anthropic directly: your requests go through AWS, which means they inherit your VPC configuration, IAM policies, CloudTrail logging, and compliance certifications.
Available Claude Models on Bedrock
Bedrock offers the full Claude model family:
- Claude Opus: The most capable model for complex reasoning and analysis
- Claude Sonnet: The balanced model for most production workloads
- Claude Haiku: The fastest and most cost-effective model for simple tasks
All models support the same features: tool use, vision, streaming, and extended thinking (where available). The API interface is slightly different from the direct Anthropic API but functionally equivalent.
Setting Up Claude on Bedrock
The setup process involves three steps:
Step 1: Enable model access. In the AWS Console, navigate to Amazon Bedrock and request access to Claude models. This requires accepting the end-user license agreement. Access is typically granted within minutes.
Step 2: Configure IAM permissions. Create an IAM policy that grants bedrock:InvokeModel and bedrock:InvokeModelWithResponseStream permissions for the specific Claude model ARNs you want to use. Attach this policy to the IAM role or user that will make API calls.
Step 3: Make API calls. Use the AWS SDK (boto3 for Python, @aws-sdk for JavaScript) to call the Bedrock Runtime API. The request format follows the Anthropic Messages API structure, wrapped in Bedrock's invocation format.
Pricing: Bedrock vs Direct API
Bedrock pricing for Claude models is generally comparable to the direct Anthropic API. Both charge per input and output token. The exact prices vary by model and region, and AWS occasionally offers different rates.
Key pricing considerations:
- On-demand pricing: Pay per token, similar to the direct API. No commitments.
- Provisioned Throughput: Reserve capacity for consistent, high-volume workloads. This guarantees throughput and can reduce per-token costs at scale.
- No data transfer costs: Requests within the same AWS region do not incur additional transfer fees.
For most workloads, the per-token cost is similar between Bedrock and the direct API. The cost advantage of Bedrock comes from infrastructure consolidation — you avoid managing a separate API key, billing account, and network configuration for Anthropic.
When Bedrock Makes Sense
Compliance requirements: If your organization requires all data processing to stay within AWS, Bedrock satisfies this without additional architecture. Data does not leave your AWS account boundary.
Existing AWS infrastructure: If you already use VPCs, IAM, CloudTrail, and other AWS services, Bedrock integrates naturally. You get logging, access control, and monitoring through tools you already manage.
Multi-model strategy: If you use multiple foundation models (Claude, Llama, Titan), Bedrock provides a single API and billing system for all of them.
Enterprise procurement: Many organizations have existing AWS enterprise agreements with committed spend. Using Bedrock lets you apply that spend to AI workloads without a separate procurement process.
Private networking: Bedrock supports VPC endpoints (PrivateLink), so your API calls never traverse the public internet. This is critical for sensitive workloads.
When the Direct API Makes Sense
Latest features first: New Claude features and models sometimes appear on the direct API before they are available on Bedrock. If you need cutting-edge capabilities immediately, the direct API may be faster.
Simpler setup: If you do not have existing AWS infrastructure, the direct Anthropic API is simpler to set up. Just get an API key and start making requests.
SDK preference: The Anthropic SDK provides a more idiomatic experience for Claude-specific features. The Bedrock API wraps the Messages API but adds AWS-specific boilerplate.
Integration with Other AWS Services
One of Bedrock's strongest advantages is integration with the AWS ecosystem:
- Lambda: Call Claude from serverless functions
- Step Functions: Orchestrate multi-step AI workflows
- S3: Process documents stored in S3 buckets
- CloudWatch: Monitor API usage, latency, and errors
- CloudTrail: Audit all model invocations
- SageMaker: Combine with custom models in ML pipelines
- EventBridge: Trigger AI workflows from events
This makes Bedrock particularly powerful for building AI-powered applications that need to integrate with existing AWS services.
Practical Tips
Start with on-demand pricing. Do not commit to provisioned throughput until you understand your usage patterns. On-demand lets you experiment freely.
Use cross-region inference. Bedrock supports invoking models in different regions. If your primary region has capacity constraints, you can route requests to another region.
Implement retry logic. Like any API, Bedrock can return throttling errors under heavy load. Implement exponential backoff in your client code.
Monitor costs closely. Token usage can be unpredictable, especially with agentic workloads. Set up CloudWatch alarms for billing thresholds.
Conclusion
Amazon Bedrock is the right choice for organizations that want to use Claude within their existing AWS infrastructure. It provides the same model capabilities with the security, compliance, and integration benefits of the AWS ecosystem.
For teams without existing AWS infrastructure or those who want the simplest possible setup, the direct Anthropic API is more straightforward. Both options give you access to the same Claude models — the choice is about infrastructure preference, not model capability.