An enterprise on AWS now has two distinct ways to reach Claude without leaving its cloud: Amazon Bedrock (AWS-operated, on AWS-managed infrastructure) and Claude Platform on AWS (Anthropic-operated, with AWS providing the authentication, IAM, and Marketplace billing layer). It's tempting to picture all of this — plus the first-party Claude API — as one big pile of Anthropic GPUs behind different doors. Anthropic's documentation says otherwise: Claude Platform on AWS uses a separate capacity pool from both the first-party Claude API and Amazon Bedrock, and notes explicitly that workloads can run on multiple platforms and fail over between them.
The isolation makes sense once you look at who runs what. On Claude Platform on AWS, the models execute on Anthropic-managed infrastructure (inference may route to Anthropic's primary cloud, and data may not reside in AWS). Claude in Amazon Bedrock runs on AWS-managed infrastructure to which Anthropic personnel have no access. These are physically and operationally different fleets, so their capacity doesn't — and can't — slosh between them.
What isolation buys you
Independent failure domains. Capacity pressure shows up as 429 rate-limit errors and, on the Claude API, 529 overloaded_error responses reflecting platform-wide load. With separate pools, a congestion event on one platform is not evidence of congestion on another: exhausting your Bedrock quota says nothing about your headroom on Claude Platform on AWS, and heavy load on the first-party API doesn't drain either AWS-side option.
Additive capacity. Because quota systems are separate too — Bedrock quotas live in AWS's per-endpoint allocations, while Claude Platform on AWS limits are managed by Anthropic's usage tiers — an organization holding allocations on both effectively holds two independent budgets for the same models. Teams that need more throughput than any single platform will grant them sometimes deliberately split steady-state traffic across platforms for exactly this reason.
A real failover story. Anthropic's docs explicitly describe running workloads on multiple platforms with failover between them. Since neither Bedrock nor Claude Platform on AWS supports the server-side fallbacks parameter, this is a client-side pattern: catch retryable errors from one platform and re-issue the request to the other.
from anthropic import AnthropicAWS, AnthropicBedrockMantle, APIStatusError
primary = AnthropicAWS() # needs AWS_REGION + ANTHROPIC_AWS_WORKSPACE_ID
backup = AnthropicBedrockMantle(aws_region="us-east-1")
def create(**kwargs):
try:
return primary.messages.create(model="claude-opus-4-8", **kwargs)
except APIStatusError as e:
if e.status_code in (429, 500, 529):
kwargs["model"] = "anthropic.claude-opus-4-8" # Bedrock ID prefix
return backup.messages.create(**kwargs)
raiseNote the model-ID translation: Bedrock prefixes IDs with anthropic., while Claude Platform on AWS uses bare first-party IDs. A failover layer has to map IDs, not just endpoints.
The fine print planners should know
Quota postures differ per pool. Isolation means you manage each side on its own terms: Bedrock through AWS quota processes (with its own two-endpoint quota split), and Claude Platform on AWS through Anthropic — where organizations stay on the Start tier until they ask for more. A failover target with a default-sized quota may absorb far less traffic than the platform it's backing up; size the backup pool for the traffic you'd actually shift.
Prompt caches don't follow you. Caching is per platform, so freshly failed-over traffic runs cold — at full input-token cost and full quota consumption — until caches rebuild on the other side.
Commercials are separate too. Anthropic documents that discounts and AWS Marketplace private offers do not transfer automatically between Bedrock and Claude Platform on AWS, and can't be applied retroactively. If both platforms are part of your resilience plan, negotiate both up front.
Compliance boundaries differ. The pools differ in more than capacity: Anthropic points organizations needing FedRAMP High, IL4/IL5, HIPAA-ready coverage, or AWS as sole data processor to Claude in Amazon Bedrock rather than Claude Platform on AWS. A failover path must clear the same review as the primary — confirm specifics with your provider.
Where to go next
See multi-platform failover patterns for the broader architecture, the platform decision framework for choosing a primary, and the platform overview for how all four 3P routes compare.