Comparison

    Claude on the Anthropic API vs Amazon Bedrock vs Vertex AI: What Changes for Policy Enforcement and Evidence

    Claude can be reached directly through Anthropic or through a cloud provider's model-as-a-service platform, Amazon Bedrock or Google Vertex AI. The access path changes the identity model, the logging surfaces, the data handling terms and the retention behavior. It does not change what a regulator asks: what the AI was allowed to do and what it actually did. Platform-native controls and logs differ by path, so the consistent answer comes from enforcing policy in the execution path and recording one evidence format for every interaction and tool call.

    Platform-independent requirements for the execution path

    • Five enforcement actions available on every interaction and tool call: allow, redact, block, rate limit, escalate for human approval.
    • Policies defined once and applied identically whether Claude is reached through the Anthropic API, Bedrock or Vertex AI.
    • Redaction applied before data reaches the model endpoint, so provider-side handling terms see less regulated content.
    • A single evidence record per interaction: identity, policy applied, decision, and outcome, in the same schema for every provider.
    • Evidence retention and residency configured by you, independent of each platform's defaults.
    • Human approval steps recorded with approver, time and result, so escalations are evidenced as well as performed.

    What actually changes between the three paths

    Amazon Bedrock and Google Vertex AI are model-as-a-service offerings: customers reach a range of models, Claude among them, through the cloud provider's platform. The Anthropic API is the separate, direct route to the same model family. The model is the same family of capability, but the surrounding system is not.

    Four things move with the path. First, identity: a cloud platform path generally relies on the cloud provider's identity and access constructs, while a direct API path relies on credentials issued for that API. Public-sector deployments show the mix in practice. One federal agency platform sends prompts to Vertex AI, Azure AI and Bedrock endpoints over TLS 1.2 or higher, and limits access through authorization, network restrictions and tenant-managed identities or API credentials. Second, logging: each platform produces its own records in its own format and scope. Third, data handling: terms on retention and training use are set per provider and per path. Fourth, residency and retention: where records live and how long they last depend on the platform and its configuration.

    For an architect, the practical consequence is that a policy written once is implemented three times, and the evidence it generates is stored three ways.

    Side-by-side: the dimensions to compare for each path

    The table lists the dimensions that matter for policy and evidence, why each varies by access path, and what to require regardless of path. Fill in the platform-specific column from your own account configuration and the provider's current documentation, and treat any gap as a requirement for the execution path.

    • Policy enforcement points: Varies: native controls are administered at different scopes (account, project, request) on each path. Require: a decision on every interaction and tool call, made before the action takes effect.
    • Identity and access control: Varies: cloud IAM on the cloud platform paths, issued credentials on the direct API path. Require: each request attributable to a user, application or agent identity in a form you control.
    • Data handling: Varies: provider terms on retention and training use differ by path. Require: redaction of regulated data before it leaves your boundary, whichever terms apply.
    • Logging and evidence: Varies: each platform has its own log surfaces, content capture options and formats. Require: one evidence schema covering policy applied, decision and outcome.
    • Retention: Varies: durations and defaults are set per platform. Require: retention of evidence set by your obligations, not by a provider default.
    • Residency: Varies: regional availability and storage location differ by path. Require: evidence stored in a region and account you choose.

    Where native controls stop

    Native controls are valuable and should be used. They are also administered by the platform, at the scope the platform defines. Account-level and project-level settings answer the question of who may call a model and from where. They do not, on their own, record that a particular prompt contained a regulated identifier, that the identifier was redacted before the model saw it, or that a particular tool call was held for human approval.

    That gap matters because the interaction and the tool call are the units reviewers care about. When Claude is used by an agent, a single user request can produce several model calls and several tool invocations. A platform log can show that calls occurred. A reviewer asking what the system was permitted to do needs the policy that applied and the decision taken at each step.

    Fragmentation is the second issue. If one team uses the Anthropic API and another uses Bedrock or Vertex AI, an auditor receives records in different formats with different fields and different retention. Reconciling them is manual work, and it happens at the point of an inquiry, when time is shortest. Until the content, retention and residency of each platform's native records are confirmed against your obligations, treat evidence completeness as an open question rather than an assumption.

    Mapping evidence to frameworks and obligations

    Frameworks such as the NIST AI RMF and the EU AI Act, and sector rules such as HIPAA, NYDFS and NAIC requirements, ask in different words for the same underlying material: controls that operate, and records showing they operated. Native platform logs may satisfy part of that. Enforcement-generated, per-interaction evidence of what was allowed, redacted or blocked is the part reviewers typically need beyond platform logs. Confirm the specific mapping against your own obligations and counsel's reading.

    Regulatory activity is also moving. As of November 2026, the GSA has proposed a clause, GSAR 552.239-7001, setting requirements for basic safeguarding of Government Data within LLM systems that process it, and the FTC has proposed a policy statement on how the Section 5 prohibition on deceptive practices applies to companies that market AI systems. Both are proposals, not final in the material reviewed. If you handle government data, identify which workloads could fall in scope, and note that a consistent per-interaction record is useful under either outcome.

    Selection criteria for any enforcement layer follow from the above. It should sit in the path rather than beside it. It should cover tool calls as well as model prompts. It should be deployable in your own cloud, so evidence stays in your account and region. And its evidence should be exportable in a form an auditor can read without learning each provider's log format.

    Where Trussed fits

    Trussed enforces policy in the execution path across models, agents, tools and MCP servers: allow, redact, block, rate limit, and escalate for human approval, at the level of individual interactions and tool calls. The enforcement itself generates continuous audit-ready evidence, mapped to HIPAA, NAIC, NYDFS, the EU AI Act and NIST. It deploys self-managed in the customer cloud, with configurable retention and data residency. It has run in continuous production in a regulated enterprise for more than 18 months and operates 100+ models under one policy plane.

    Trussed does not replace the native controls of the Anthropic API, Bedrock or Vertex AI. It adds the consistent enforcement point and the single evidence record that those controls do not provide across paths.

    One policy decision point, three access paths

    Anthropic API

    Direct access to Claude through Anthropic.

    Amazon Bedrock

    Claude reached through Amazon's model-as-a-service platform.

    Google Vertex AI

    Claude reached through Google's model-as-a-service platform.

    Execution path

    Allow, redact, block, rate limit or escalate, with one evidence record per interaction.

    Apply one policy and one evidence format to every Claude path

    See how enforcement in the execution path produces consistent, audit-ready evidence whether Claude is accessed through the Anthropic API, Amazon Bedrock or Vertex AI.

    Request a Demo