Upgrading from Legacy Bedrock to the Current Mantle Surface

Amazon Bedrock serves Claude through two surfaces. The legacy one — InvokeModel and Converse on the bedrock-runtime endpoint — wraps Claude in AWS API conventions: ARN-versioned model identifiers, a "anthropic_version": "bedrock-2023-05-31" body field, and AWS event-stream encoding for streaming. The current surface, "Claude in Amazon Bedrock," serves Anthropic's Messages API directly at https://bedrock-mantle.{region}.api.aws/anthropic/v1/messages with the same request body shape as the first-party API and standard SSE streaming. The legacy surface remains available, but current-generation models are pushing the decision: Claude Sonnet 5 is not on the legacy surface at all, and Fable 5, Opus 4.8, and Opus 4.7 have no ARN-versioned IDs.

The code changes

With the SDK, the visible diff is small:

from anthropic import AnthropicBedrockMantle

client = AnthropicBedrockMantle(aws_region="us-east-1")
msg = client.messages.create(
    model="anthropic.claude-opus-4-8",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello"}],
)

Credentials keep working unchanged: the client resolves them via the standard AWS precedence — constructor args, env vars, config files, SSO, roles, instance metadata. Note the ID form changes twice: the routing prefix (global./us.) disappears, and 4.6+ models go dateless. Routing is now chosen by endpoint configuration — the current surface offers global endpoints (no premium) and regional endpoints (10% premium, for data residency) — rather than by editing the model ID string. One behavioral note if you relied on it: the legacy Converse API needed citations enabled for full visual PDF understanding; check your PDF paths as part of validation.

The operational changes people miss

• IAM policies must change. The Mantle surface uses a distinct action namespace: grants of bedrock:InvokeModel do nothing for it. Principals need bedrock-mantle:CreateInference on the allowed model ARNs, and VPC users need the separate com.amazonaws.{region}.bedrock-mantle PrivateLink endpoint.
• Quotas are separate buckets. The two endpoints track per-model quotas independently, even for the same model — your legacy-surface quota history does not carry over. Mantle enforces input-TPM and output-TPM (no RPM quotas), reserves input tokens plus max_tokens at admission, and does not count cached reads against input TPM. Increases go through an AWS Support case, not the Service Quotas console.
• Model invocation logging does not follow you. Bedrock's invocation logging captures only bedrock-runtime calls; Mantle-endpoint calls are not captured by it. Claude in Amazon Bedrock emits to CloudWatch and CloudTrail — rebuild your audit pipeline against those before cutting over, not after.
• The FTU form is not required. The one-time Anthropic use-case form that legacy Bedrock accounts submit does not apply to models accessed through the bedrock-mantle endpoint — one less onboarding step for new accounts.
• Beta features become reachable. The legacy surface does not support the anthropic-beta header; see the beta-header article for what that unlocks and what still is not on Bedrock at all (Anthropic's Message Batches API, Files API, and server-side web tools remain unavailable on either surface).

Where to go next

For the ID mechanics in detail, read the anthropic. prefix rule. If you are weighing Bedrock against Anthropic-operated options on AWS, the legacy API comparison and the Claude Platform on AWS migration guide frame the alternatives.

Sources

• Claude in Amazon Bedrock
• Claude on Amazon Bedrock (legacy)
• Bedrock Mantle endpoint quotas (AWS)
• Model invocation logging (AWS)
• Bedrock model access (AWS)