AWS Publishes a Claude Code Route for FedRAMP Claude 5.5 in GovCloud

AWS published the GovCloud Claude Code route on October 5

Amazon Web Services published a deployment guide on October 5, 2026 for running Claude Code against Amazon Bedrock in AWS GovCloud (US). The event is new implementation guidance for regulated AI-assisted development, not a new foundation model: AWS had already announced Opus 5.5 in GovCloud on September 22 and Sonnet 5.5 there on September 28.

The guide is explicit that its approach may not fit every organization or compliance program. That qualification matters. A model appearing in GovCloud and carrying a service authorization label does not make a developer workstation, repository, shell command, MCP server or CI runner compliant by inheritance.

Runtime and Mantle expose different control surfaces

The bedrock-runtime endpoint uses InvokeModel and Converse. AWS documents it in GovCloud US-West and US-East with Bedrock Guardrails, Knowledge Bases, Agents and invocation logging. The bedrock-mantle endpoint uses the native Anthropic Messages API and adds server-side tools, background inference and Projects, but AWS lists it only in GovCloud US-West.

Guardrails and invocation logging are exclusive to bedrock-runtime in the reviewed guide. Mantle also uses a separate bedrock-mantle IAM namespace; bedrock permissions do not cover it. Treat an endpoint switch as a control and API migration, not a transport toggle. Reapprove the Region, IAM actions, logging evidence, filtering behavior, request schema and fallback path.

Pin the us-gov profile instead of trusting an alias

AWS’s manual setup uses us-gov.anthropic.claude-sonnet-5-5 or us-gov.anthropic.claude-opus-5-5 in ANTHROPIC_MODEL. The wizard path selects Amazon Bedrock, authentication, us-gov-west-1 and model pins. The runtime route requires at least bedrock:InvokeModel, bedrock:InvokeModelWithResponseStream, bedrock:ListInferenceProfiles and bedrock:GetInferenceProfile; Mantle requires its own actions.

Anthropic documents startup access checks and fallback behavior when a default is unavailable. AWS also says Claude Code currently defaults to Opus 5.5 as its primary model. An enterprise baseline should therefore pin full model IDs, constrain allowed models centrally and fail closed when the approved profile is unavailable rather than silently crossing a capability, certification or price boundary.

FedRAMP Class D does not mean every DoD tier is covered

AWS’s certification matrix, last updated October 5, lists Claude 5.5 Opus and Sonnet under the GovCloud FedRAMP Class D column. The same two entries are not marked in the separate DoD CSP SRG IL4/IL5 column. The deployment guide consequently points workloads that require IL4/IL5 authorization to Claude Sonnet 5, not Sonnet 5.5 or Opus 5.5.

Use the live matrix, account evidence and applicable authorization package at decision time. Do not collapse FedRAMP Class D, ITAR handling and DoD impact levels into one label. The AWS GovCloud documentation also says certain customer-defined metadata may leave GovCloud when the customer asks AWS to investigate an issue, and forbids export-controlled data in model-evaluation metadata.

Public numeric GovCloud pricing was not stated

The October 5 guide says Opus 5.5 has a higher per-token cost than Sonnet 5.5 and recommends Sonnet as a default for many teams, but it does not quote numeric GovCloud rates. The September 28 Sonnet availability notice also omits token prices. AccessAllGPT is therefore not importing direct Claude API prices or a commercial-Region quote into this GovCloud route.

Capture the live Bedrock rate for the exact account, profile, endpoint and Region before approval. Forecast uncached input, cache writes and reads, output, retries, long agent trajectories and surrounding logging or guardrail services. Add per-user token limits and alerts, and verify the bill after a bounded pilot. “FedRAMP” is an authorization attribute, not a price plan.

Claude Code remains a privileged local agent

Claude Code can read and edit repositories, execute tests and shell commands, inspect Git history, create commits and pull requests, connect to MCP tools and run in CI. Bedrock protects the inference route; it does not confine those local capabilities. AWS’s own guide says the software running on developer machines requires a separate security assessment.

Use IAM Identity Center and temporary role credentials, centrally managed permissions, repository-scoped execution, approved MCP servers, hooks that enforce deterministic policy, egress controls and complete local plus Bedrock audit trails. Keep consequential writes behind external authorization. Test whether logs contain source, prompts, tool arguments or secrets before enabling them broadly.

What regulated teams should do next

Adopt a bedrock-runtime pilot when the required model is certified for the workload, Guardrails and invocation logging are part of the evidence plan, and one GovCloud Region is approved. Constrain Claude Code to read-only analysis or patch proposals until command, repository and MCP policies pass. Wait when the price, authorization package, model access or metadata path is unresolved. Reject silent model fallback or a Mantle migration that drops required logging controls.

Start with a non-production repository and synthetic non-sensitive tasks. Record the exact Claude Code version, full model profile, endpoint, Region, IAM role, permissions, settings policy, tools, logs, quotas, token ceiling, rate capture and stop conditions. Exercise denied commands, unavailable-model behavior, fallback blocking, log retrieval and credential expiry before allowing regulated source or autonomous CI execution.

Disclosures

AccessAllGPT received no AWS or Anthropic account, credentials, credits, early information, briefing, demo, authorization evidence, review or compensation for this article. Neither company sponsored, reviewed or endorsed it. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with Amazon Web Services, Anthropic or agencies and standards bodies cited. Publication-wide relationships are listed on the disclosures page.