Claude Platform on AWS vs Amazon Bedrock: Which Should You Use in 2026?

If you're an AWS developer trying to add Claude to your stack, you now have three options — and the newest one launched just weeks ago. AWS teams have historically defaulted to Amazon Bedrock for Claude access. Now Anthropic has changed the game with Claude Platform on AWS, a fundamentally different approach to running Claude inside your AWS account.

This guide breaks down exactly what's different, when each option wins, and how to set up Claude Platform on AWS from scratch — including a decision framework you can use in architecture reviews today.

The Three Ways to Run Claude on AWS

Before comparing, let's clarify the landscape. As of mid-2026, you have three distinct paths:

The key insight: Claude Platform on AWS exists to close the feature gap between Bedrock and the native API while letting you keep AWS authentication, billing commitments, and audit tooling.

Claude Platform on AWS was announced at general availability in May 2026, following Anthropic and Amazon expanding their partnership with a commitment of more than $100 billion over ten years in compute infrastructure, including Graviton and Trainium4 chips. The timing is not coincidental — this is Anthropic's answer to enterprise AWS teams who want native features without leaving the AWS ecosystem.

What Claude Platform on AWS Actually Gives You

When you activate Claude Platform on AWS, you get access to the full Anthropic native platform surface — the same feature set that ships to the direct API on day one:

• Claude Managed Agents (beta) — multi-step autonomous agents with memory and tool use
• Agent Skills — reusable capability modules you attach to agents
• MCP Connectors — connect Claude to any Model Context Protocol server
• Code Execution — sandboxed Python runtime inside agent sessions
• Files API (beta) — upload and reference documents in agent sessions
• Prompt Caching — cache prefixes for up to 5 minutes, dramatically cutting costs on repeated context
• Batch Processing — async bulk inference at reduced pricing
• Web Search & Fetch — grounded, citation-aware web access
• Claude Console — the full development and evaluation UI

The authentication and billing story is equally clean. You authenticate with AWS SigV4 — the same mechanism your existing AWS SDKs use. Audit logs flow to CloudTrail. Billing consolidates into your existing AWS invoice and retires against your Enterprise Discount Program or AWS Committed Use commitments.

No new vendor relationships. No separate billing portal. No context-switching between consoles.

Claude Platform on AWS vs Amazon Bedrock: The Real Differences

Feature Availability

This is the biggest decision factor for most teams. Amazon Bedrock runs Claude models excellently, but Anthropic ships features to its own API surface first. Features like Managed Agents, Skills, the Files API, and native MCP connectors either land on Bedrock months later or arrive with reduced fidelity.

Claude Platform on AWS solves this by running Anthropic-managed infrastructure — you're calling Anthropic's stack, just authenticated and billed through AWS.

If your application architecture requires Managed Agents, Skills, or code execution as core building blocks, Claude Platform on AWS is the right call. Waiting for Bedrock feature parity on agentic capabilities can mean months of lag on a fast-moving product.

Data Residency and Security Boundary

This is where Bedrock wins for specific compliance requirements. Claude on Amazon Bedrock runs inside the AWS security boundary — your prompts, completions, and session data never leave AWS infrastructure. That's crucial for workloads in regulated industries with strict data residency requirements (HIPAA, FedRAMP, certain financial compliance frameworks).

Claude Platform on AWS is operated by Anthropic. Data is processed outside the AWS security boundary, on Anthropic's infrastructure. Anthropic has its own strong security posture and compliance certifications, but if your security review specifically requires "data stays within our AWS account boundary," Bedrock is the correct choice.

Billing Model

Bedrock charges per-token with standard AWS per-service line items. Claude Platform on AWS introduces a new billing unit — Claude Consumption Units (CCUs) — which map to model usage and retire against your AWS Marketplace spend.

For teams with large AWS committed spend, this can be a meaningful advantage: your Claude usage now directly contributes to meeting existing commitments. For teams without pre-committed AWS spend, the billing models are roughly comparable in effective cost.

Multi-Model Pipelines

If your application routes across multiple foundation model providers — Claude for reasoning, Titan for embeddings, Stable Diffusion for images — Bedrock's unified inference layer is architecturally cleaner. All models in one API surface, one billing model, one place for Guardrails configuration.

Claude Platform on AWS is Anthropic-only. If you're already all-in on Claude (common for applications targeting CCA-level capabilities), this isn't a constraint. But if you need a multi-model pipeline, Bedrock's breadth is a genuine advantage.

When to Choose Claude Platform on AWS

Getting Started with Claude Platform on AWS

Setting up Claude Platform on AWS takes about 15 minutes for an AWS team already familiar with IAM.

Step 1: Subscribe via AWS Marketplace

Navigate to the Claude Platform on AWS listing in AWS Marketplace and subscribe. This links your AWS account to Anthropic's platform.

Step 2: Configure IAM Permissions

Create an IAM role or policy that grants claude-platform:InvokeModel permissions. Claude Platform on AWS uses SigV4 signing, so your existing AWS credential chain works:

pythonimport boto3
import anthropic
from anthropic import AnthropicBedrock

# Using the Claude Platform on AWS endpoint
client = anthropic.Anthropic(
    base_url="https://api.claude-platform.aws.anthropic.com",
    auth_token=None  # SigV4 handles auth automatically
)

# Or use the AWS-native SDK wrapper
session = boto3.Session()
credentials = session.get_credentials()

message = client.messages.create(
    model="claude-opus-4-7-20260501",
    max_tokens=1024,
    messages=[
        {"role": "user", "content": "Explain prompt caching to an AWS developer."}
    ]
)

Step 3: Enable Managed Agents (Optional)

If you're building agentic applications, enable Managed Agents from the Claude Console. You can access the console at console.claude-platform.aws.anthropic.com — it's the full Anthropic development UI, surfaced inside your AWS context.

Step 4: Configure CloudTrail Logging

Audit events from Claude Platform on AWS appear in CloudTrail under the claude-platform.amazonaws.com event source. Set up a trail if you don't have one, and verify events are flowing before deploying to production.

Step 5: Test with a Managed Agent

Here's a minimal agent that uses code execution — a feature not natively available on Bedrock without custom Lambda integrations:

python# Create a managed agent session with code execution enabled
agent_session = client.agents.sessions.create(
    agent_id="your-agent-id",
    tools=["code_execution", "web_search"]
)

# Run a task
result = client.agents.sessions.run(
    session_id=agent_session.id,
    input="Analyze this CSV data and produce a summary with key statistics",
    files=["s3://your-bucket/data.csv"]
)

print(result.output)

Pricing: What to Expect

Claude Platform on AWS uses the same underlying model pricing as the direct Anthropic API, denominated in Claude Consumption Units:

• Claude Opus 4.7: ~$15 / 1M input tokens, ~$75 / 1M output tokens
• Claude Sonnet 4.6: ~$3 / 1M input tokens, ~$15 / 1M output tokens
• Prompt Caching: 90% cost reduction on cached prefixes

The billing innovation is that these charges appear on your AWS invoice and retire against your AWS Marketplace committed spend. For organizations with large AWS commitments, this can be financially significant.

Note: Bedrock charges are structurally similar but differ in unit denomination. Run a cost model against your expected token volume before committing — both options are competitive, with the real differentiation in features and data residency rather than raw price.

The CCA Exam Angle: What This Means for Claude Certified Architects

For developers pursuing the Claude Certified Architect (CCA-F) certification, Claude Platform on AWS introduces new architecture patterns that are highly likely to appear in exam scenarios:

• Understanding when to choose Bedrock vs. Claude Platform on AWS is a core architectural decision pattern
• SigV4 authentication and IAM integration with Claude is a testable integration pattern
• Managed Agents with code execution and Skills represent the agentic architecture tier that CCA exams emphasize
• Data residency tradeoffs in enterprise Claude deployments are a recurring theme in architecture questions

If you're studying for CCA, make sure your mental model includes all three Claude-on-AWS options (Bedrock, Claude Platform on AWS, direct API) and the conditions under which each is architecturally correct.

Key Takeaways

• Claude Platform on AWS is Anthropic-operated, not AWS-operated — data leaves the AWS security boundary
• Feature parity is instant — every new native Claude capability ships to Claude Platform on AWS on day one, no Bedrock lag
• Billing consolidates — usage appears on your AWS invoice and retires against Marketplace commitments
• Bedrock still wins for compliance-sensitive workloads requiring data inside the AWS boundary, multi-model pipelines, and PrivateLink
• Claude Platform on AWS wins for agentic applications, MCP-connected workflows, and teams that need today's features today

Next Steps

Ready to test Claude Platform on AWS in your stack? Start with the official AWS announcement and the architecture decision guide from Caylent for a detailed decision tree.

If you're preparing for the Claude Certified Architect exam, these architecture tradeoffs — Bedrock vs. Claude Platform vs. direct API — are exactly the kind of decision patterns CCA tests you on. Try our free CCA practice questions to see how exam scenarios frame Claude-on-AWS architectural choices, then unlock the full 300-question bank when you're ready to certify.

For a broader view of the Claude ecosystem changes shipping this month, our guides on Claude Managed Agents and building MCP servers for Claude are the best next reads.