Claude Platform on AWS: The Third Deployment Path Every Operator Routing Policy Must Now Account For
Claude Platform on AWS gives teams Anthropic-managed inference through AWS billing — full beta headers, Agent Skills, and a separate capacity pool that unlocks a new multi-platform failover strategy.
Archive item produced with AI assistance from the cited source and published without individual review. Editor of record: Joe Werner.
When Claude Code 2.1.198 shipped the anthropicAws gateway provider last week, it surfaced a deployment option that many teams have been treating as a footnote: Claude Platform on AWS. This is not Amazon Bedrock with a different label. It is a structurally different path — and engineering teams that have not mapped it into their provider routing policy are making that decision by default rather than by intent.
What Claude Platform on AWS actually is
The surface-level description is deceptively simple: Claude on AWS with Anthropic-managed inference and AWS Marketplace billing. The architectural difference from Amazon Bedrock runs deeper than that.
On Amazon Bedrock, AWS operates the inference stack. Anthropic personnel have no access to the infrastructure. AWS is the data processor, the compliance boundary owner, and the capacity manager. That is the correct choice when FedRAMP High, IL4, IL5, HIPAA-ready compliance, or strict AWS-operated data residency is required.
On Claude Platform on AWS, Anthropic operates the inference stack. AWS provides authentication (SigV4 or API key), IAM-based access control, and billing integration through the Marketplace. Anthropic is the data processor. The inference data stays under Anthropic's operational control — but the procurement, identity, and audit surface runs through AWS.
The practical consequence for operators: Claude Platform on AWS delivers the same API surface as the first-party Claude API, with the same cadence for new features and beta headers.
The API surface gap that changes routing decisions
This is the part that most comparison articles undersell. Here is what the difference means in practice:
Beta headers work. The anthropic-beta header passes through on Claude Platform on AWS. On Amazon Bedrock (in its current generation), anthropic-beta is not supported. Any team using mcp-tunnels-2026-06-22, managed-agents-2026-04-01, fallback-credit-2026-06-01, server-side-fallback-2026-06-01, or any other recent Anthropic beta feature cannot access those features through Amazon Bedrock today. They can access them through Claude Platform on AWS.
Agent Skills are available. Claude Managed Agents with Agent Skills — the hosted skill execution environment — runs on Claude Platform on AWS. It does not run on Amazon Bedrock (which requires code execution as an alternative path) or on the legacy Bedrock integration.
Feature availability lag is eliminated. Bedrock releases track Anthropic's schedule but are not same-day. Claude Platform on AWS typically matches the first-party API release cadence on the same day a feature ships.
These three differences determine which teams can even use Claude Platform on AWS for a given workload — and they directly affect routing policy for any team running a multi-provider setup.
Why it matters for AI engineering teams
The most direct implication: if your team has standardized on AWS procurement and is hitting beta-feature access limitations on Bedrock, Claude Platform on AWS is the resolution path that does not require moving off AWS billing or re-integrating with a new identity system.
Two operator scenarios where this changes the routing decision:
Scenario 1: Teams using Claude Managed Agents or Anthropic beta features through AWS. If your organization has already signed an AWS Marketplace subscription for AI services and wants to run Claude Managed Agents with per-session config overrides, event deltas, or deployment webhooks (all shipped June 30), those features require the Claude API surface. Claude Platform on AWS delivers that through AWS billing without requiring a separate Anthropic direct contract.
Scenario 2: Teams building a multi-platform fallback strategy. Claude Platform on AWS runs on a separate capacity pool from both the first-party Claude API and Amazon Bedrock. Anthropic explicitly supports running workloads on more than one platform and failing over between them. This means a team can configure a fallover chain: primary on Anthropic direct API → overflow to Claude Platform on AWS → overflow to Claude in Amazon Bedrock (for supported models). Each leg uses a different capacity pool, which is the operational point: a capacity event on one path does not necessarily affect the others.
The router/operator angle
Three-path Claude routing requires explicit policy. Most routing configurations today treat "Anthropic" as a single provider. Claude Platform on AWS introduces a third distinct base URL (aws-external-anthropic.{region}.api.aws), a separate SDK client class (AnthropicAWS in Python beta), and a separate capacity pool. Each of these is a distinct upstream in a gateway's provider registry. Leaving them unconfigured means leaving the fallback chain incomplete.
The key configuration decision for each path:
| Path | Base URL pattern | Capacity pool | Beta headers | When to use |
|---|---|---|---|---|
| Anthropic direct | api.anthropic.com | Anthropic primary | Yes | First-party billing, broadest feature coverage |
| Claude Platform on AWS | aws-external-anthropic.{region}.api.aws | Separate Anthropic | Yes | AWS billing + Anthropic features, multi-pool fallback |
| Claude in Amazon Bedrock | bedrock-mantle.{region}.api.aws | AWS | No | AWS-operated compliance boundary, no anthropic-beta |
Data residency is per-request, not per-account. On Claude Platform on AWS, inference may route to Anthropic's primary cloud by default. The inference_geo parameter lets teams pin specific requests to a geography — us, eu, or other supported values. This is a request-level signal, not a region-level configuration. Teams with data residency requirements need to set this per request, not rely on the AWS region selection alone.
The AWS PrivateLink path is available. For teams that require private VPC connectivity to Claude without traversing the public internet, Claude Platform on AWS supports AWS PrivateLink. This is the same pattern available for Amazon Bedrock but on the Anthropic-operated endpoint. It matters for enterprise deployments where egress to public endpoints must be minimized.
What TheRouter users should watch or try
The anthropicAws provider added to Claude Code 2.1.198 is the same architectural path. Teams testing Claude Code with anthropicAws configured as a gateway failover are already using this capacity pool. If that configuration is working reliably in Claude Code, the same underlying endpoint is available for broader API workloads.
For teams evaluating provider routing policy on TheRouter, the three-path model changes the upstream configuration from a two-option (Anthropic vs Bedrock) to a three-option choice. Each path requires its own authentication configuration (SigV4 for Bedrock paths; API key or SigV4 for Claude Platform on AWS; API key for Anthropic direct), which affects how credential rotation and access control are managed at the gateway layer.
The separate capacity pool property is the most operationally significant detail for routing teams. When one pool is under load-induced latency, the others may not be. Building a failover chain that spans pools — not just providers — is the routing strategy this architecture enables.
Anthropic Inference Hooks Put a Pre-Inference Gate at the Provider Layer: What It Means for Your Routing Architecture
Anthropic's new Inference Hooks let enterprise organizations intercept every governed Claude prompt before the model runs. For teams already filtering at the gateway layer, this creates a dual-gate architecture that changes where enforcement belongs.
Claude Text Watermarking Is Coming to All Operators: What the SynthID-Text Detection API Means for Your Pipeline
Future Claude models will embed a SynthID-Text watermark globally, with no opt-out. A detection API is coming. Operators running content-review or compliance pipelines need to understand what the watermark proves — and what pipeline modifications can silently degrade it.
Fable 5 Biology Classifier Fix: The Silent Model-Swap Your API Billing Never Warned You About
Fable 5 biology classifier false-positive fix cuts fallbacks 85%. For API operators, this exposed a silent billing risk: Fable 5 requests were being served by Opus 5 without warning. Here is what to audit before assuming model parity returns.