Deploy Claude on the cloud you already trust.

The basics

What does "3P" mean?

Anthropic (the company behind Claude) offers two ways to use its models. First-party (1P) means calling Anthropic's own API directly with an Anthropic account. Third-party (3P) means using the exact same Claude models through a cloud platform your company already uses. Same models, same quality — different billing, contracts, and access controls.

The business case

Why enterprises choose the 3P route

Step 1 · Choose

The four ways to run Claude in the enterprise

Pick based on where your workloads already live. All four serve the same Claude models through the same Messages API shape — switching later is mostly a one-line client change.

Step 2 · Connect

Quickstart: your first Claude call on each platform

All examples use the official Anthropic Python SDK (pip install anthropic) — each platform has a dedicated client class, and after construction the API is identical. TypeScript, Java, Go, C#, Ruby, and PHP SDKs follow the same pattern.

# pip install -U "anthropic[aws]"
from anthropic import AnthropicAWS

client = AnthropicAWS()  # reads AWS_REGION + ANTHROPIC_AWS_WORKSPACE_ID

response = client.messages.create(
    model="claude-opus-4-8",          # bare model ID — no prefix
    max_tokens=1024,
    messages=[{"role": "user", "content": "Summarize this contract clause: ..."}],
)
print(response.content[0].text)

# pip install -U "anthropic[bedrock]"
from anthropic import AnthropicBedrockMantle

client = AnthropicBedrockMantle(aws_region="us-east-1")

response = client.messages.create(
    model="anthropic.claude-opus-4-8",   # note the anthropic. prefix
    max_tokens=1024,
    messages=[{"role": "user", "content": "Draft a reply to this customer email: ..."}],
)
print(response.content[0].text)

# pip install -U "anthropic[vertex]"
from anthropic import AnthropicVertex

client = AnthropicVertex(project_id="my-gcp-project", region="global")

response = client.messages.create(
    model="claude-opus-4-8",          # bare model ID — no prefix
    max_tokens=1024,
    messages=[{"role": "user", "content": "Classify this support ticket: ..."}],
)
print(response.content[0].text)

# pip install -U anthropic
from anthropic import AnthropicFoundry

client = AnthropicFoundry(
    api_key="<your-foundry-key>",     # from the Azure portal
    resource="my-foundry-resource",
)

response = client.messages.create(
    model="claude-opus-4-8",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Extract the key dates from this memo: ..."}],
)
print(response.content[0].text)

Step 3 · Plan

Feature availability at a glance

The core Messages API — chat, streaming, tool use, vision, PDFs, prompt caching — works on every platform. Differences show up in the newer, agent-oriented surface. Check this before you commit to an architecture.

✔ = generally available · β = beta · ✘ = not supported. Snapshot as of July 2026 — always confirm against the official platform availability table before finalizing an architecture.

Step 4 · Size

Which model for which job

Prices are Anthropic list prices per million tokens; on 3P platforms your effective rate is set by the cloud provider and may draw down committed-use discounts.

Remember: on Amazon Bedrock, prepend anthropic. to every ID above.

Step 5 · Ship

Production deployment checklist

The eight things to have in place before your first production workload.

• Least-privilege accessCreate a dedicated IAM role / service account per application. Never share one credential across teams.
• Region & data residencyPin inference to regions that satisfy your compliance requirements; confirm with legal where request data is processed.
• Quotas & rate limitsCheck the default tokens-per-minute quota for your account tier and request increases before launch, not during it.
• Prompt caching from day onePut stable content (system prompt, tool definitions, reference docs) first and cache it — typical savings on repeated context are up to 90% of input cost.
• Retry & fallback logicThe official SDKs retry rate limits and server errors automatically. Add graceful degradation (e.g. fall back from Opus to Sonnet) for peak load.
• Cost monitoringTag Claude usage with your cloud cost-allocation tags and set budget alerts. Track cache hit rate — a low one usually means a prompt-structure bug.
• ObservabilityLog request IDs, token usage, and stop reasons for every call. You'll need them for support tickets and cost analysis.
• Model version pinningReference models by their documented IDs and test before upgrading — new model generations can change behavior and token counts.

Common questions

FAQ