Advanced

AWS Bedrock (Claude)

AWS Bedrock is Amazon's managed AI service that hosts Claude (and other foundation models) within the AWS infrastructure. Instead of calling the Anthropic API directly, you call Bedrock — and Claude runs inside AWS. This matters for organisations that require AWS billing, VPC isolation, IAM authentication, or data that must never leave a specific AWS region.

Why Use Bedrock Instead of Direct API?

Reasons to use Bedrock

AWS billing: Consolidate AI spend into your existing AWS account / AWS contract — use EDP credits, Reserved Capacity, or RI agreements
IAM authentication: No separate API key to manage — use IAM roles and AWS Signature v4 auth, consistent with your other AWS services
VPC deployment: Invoke Claude via a VPC endpoint — requests never traverse the public internet
Data residency: Data stays within the AWS region you select — critical for EU data sovereignty, HIPAA, and FedRAMP-adjacent requirements
AWS CloudWatch integration: Logs and metrics flow automatically to your existing observability stack
Guardrails and Model Evaluation: Bedrock-native features wrap Claude with additional safety and evaluation tooling

Reasons to use direct Anthropic API

You want the latest model versions immediately (Bedrock has a lag on new releases)
You prefer simpler authentication (API key vs IAM)
You're building outside AWS or want cloud-agnostic code
Direct API pricing may be lower for some usage patterns (compare before committing)
You need features Bedrock hasn't yet added (e.g. prompt caching, newer API capabilities)

Enabling Claude Models on Bedrock

Claude models on Bedrock require explicit model access requests (unlike some AWS services that are available by default):

Open the AWS Console → Amazon Bedrock → Model access
Find the Claude model(s) you need (e.g. Claude Sonnet 4.6, Claude Haiku 4.5)
Click "Request access" — for most Claude models this is approved immediately
Once approved, the model status shows "Access granted"
Repeat in each AWS region where you need access — access is per-region

Model availability lags direct API

New Claude models typically appear on the direct Anthropic API first, then on Bedrock weeks to months later. If you need the latest Claude version immediately, the direct API is the only option. Check the Bedrock model catalogue for the current list of available Claude versions.

API Differences: Converse vs Messages

Bedrock exposes Claude through two different API shapes — choose based on your needs:

Bedrock Converse API

AWS-unified interface that works across all Bedrock models. Same request shape regardless of model provider.

Model-agnostic — switch between Claude, Titan, Llama without changing code
Supported in all AWS SDKs (boto3, AWS SDK for JS, etc.)
Slightly less feature-rich — may not expose all Anthropic-specific parameters

InvokeModel with Anthropic format

Passes the Anthropic Messages API request body directly through Bedrock.

Same request shape as the direct Anthropic API
Easier to migrate existing Anthropic SDK code to Bedrock
Access to Anthropic-specific parameters (extended thinking, etc.)
Can use the Anthropic SDK with a custom base URL pointing to Bedrock

Regional Availability

Claude on Bedrock is available in multiple AWS regions — US East (N. Virginia), US West (Oregon), EU (Frankfurt, Ireland), Asia Pacific (Tokyo, Singapore, Sydney) and others. Key considerations:

Not all Claude model versions are available in all regions — check the Bedrock documentation for region-model availability matrix
Latency varies significantly by region — test from your primary traffic origin
For GDPR compliance, data must stay in EU regions (Frankfurt, Ireland) — do not route EU user data through US regions
Cross-region inference is available — Bedrock can automatically route to the nearest available region, but data may leave your primary region

Cost Comparison

Bedrock pricing for Claude is per-token (input + output), similar to direct API. Differences:

Bedrock list prices are typically slightly higher than direct Anthropic API list prices
If you have an AWS Enterprise Discount Program (EDP) or committed use, those discounts apply to Bedrock spend — potentially making it cheaper in aggregate
Bedrock charges through your AWS bill — apply cost allocation tags to track AI spend alongside other AWS infrastructure costs
Provisioned Throughput on Bedrock allows reserving model capacity for a fixed hourly cost — useful for predictable, high-volume workloads

Checklist: Do You Understand This?

Bedrock hosts Claude within AWS — use it for IAM auth, VPC isolation, AWS billing consolidation, and data residency requirements
Claude models require explicit access requests per region in the Bedrock console
New Claude versions appear on direct API first — Bedrock availability lags by weeks to months
Converse API is model-agnostic; InvokeModel with Anthropic format preserves full Anthropic API feature access
For GDPR, pin traffic to EU regions; for HIPAA, combine Bedrock with a BAA from AWS