Where should you run Claude: Anthropic's API, Bedrock or Google Cloud? Use Anthropic's Claude API when you want every feature as soon as it ships — Batches, Files, server tools, Managed Agents. Use Claude in Amazon Bedrock or Claude on Google Cloud when Claude must run inside that cloud, under its identity, data handling and billing, and you can live without the features those platforms do not support. Consider Claude Platform on AWS if you want Anthropic-operated Claude billed through AWS Marketplace. Running two of them gives you separate quotas and the safest failover for Claude-specific workloads: the same model on another platform.
The platforms side by side
| Claude API | Claude in Amazon Bedrock | Claude on Google Cloud | Claude Platform on AWS | |
|---|---|---|---|---|
| Operated by | Anthropic | AWS | Google Cloud | Anthropic |
| Billing | Anthropic | AWS | Google Cloud | AWS Marketplace (CCUs) |
| API shape | Messages API | Messages API at /anthropic/v1/messages | Messages API; model in URL, anthropic_version in body | Claude API |
| Model id example | claude-opus-5-5 | anthropic.claude-opus-5-5 | claude-opus-5-5 in the endpoint path | Claude API ids |
| Regional option | inference_geo: "us" at 1.1x | Regional endpoints, +10% | Multi-region (us, eu) and regional, +10% | inference_geo: "us" at 1.1x |
| Message Batches, Files API | Yes | No | No | See Anthropic's docs |
| Code execution, MCP connector | Yes | No | No | See Anthropic's docs |
| Web search | Yes | No | Yes | See Anthropic's docs |
Features: the gap that decides most cases
Anthropic lists what each cloud does not support.
Claude in Amazon Bedrock supports the Messages API, prompt caching, thinking, tool use (including bash, computer use, memory and text editor tools), citations and structured outputs. It does not support URL sources for images and documents or the Files API; server-side tools (code execution, web search, web fetch, advisor); Agent Skills, the MCP connector or programmatic tool calling; the Message Batches, Models, Admin, Compliance and Usage and Cost endpoints; Claude Managed Agents; server-side fallback; or the newest computer use and browser use toolsets1.
Claude on Google Cloud supports the Messages API, prompt caching, thinking, tool use including browser and computer use, the web search tool, citations and structured outputs. It does not support URL sources or the Files API; code execution, web fetch or advisor; Agent Skills, the MCP connector or programmatic tool calling; the Message Batches, Models, Admin, Compliance and Usage and Cost endpoints; Managed Agents; or server-side fallback2.
Two consequences are easy to underestimate:
- No Message Batches means no half-price batch processing through Anthropic's Batch API on those clouds; plan background workloads differently there.
- No Usage and Cost API means your Claude spend on a cloud is visible through that cloud's billing tools, not Anthropic's, so cost reporting needs a second source.
For Anthropic-operated access on AWS, Anthropic points to Claude Platform on AWS, with AWS Marketplace billing and "typically same-day feature access"1. It bills in Claude Consumption Units at a fixed $0.01 each, converted from token usage at standard rates after any negotiated discount3.
Where inference runs, and what it costs
Bedrock offers global endpoints, which route across regions with no premium, and regional endpoints for data residency, which carry a 10% premium; inference profiles route within a geography (US, EU, JP or AU)1. Google Cloud offers global endpoints with no premium, multi-region endpoints within the US or EU, and regional endpoints, the last two with a 10% premium; provisioned throughput requires a regional endpoint, and specific regional endpoints serve Claude Sonnet 4.6 and earlier while newer models use global or multi-region endpoints2.
On Anthropic's own API, US-only inference with inference_geo: "us" applies a 1.1x multiplier to every token category on Claude 4.6 and later3. The list rates on Bedrock and Google Cloud are published by those platforms; Anthropic's pricing page points to them rather than restating them3. Claude API pricing has the first-party rate card.
Quotas and access
On Bedrock, Anthropic documents a default quota of 2 million input tokens per minute, with up to 5 million input and 500,000 output tokens per minute available without additional Anthropic approval; AWS enforces requests-per-minute limits on its side. Access is set per model by Bedrock: some models are open to all Bedrock customers, others need checking in the model access console1. On the Claude API, limits come from your usage tier.
Because each platform has its own quota, adding a second platform for the same model is the cleanest way to raise Claude capacity before a launch. Load balancing LLM API keys explains why more keys on one platform do not.
Data handling and logging
On Bedrock, data handling is governed by Amazon Bedrock, and Anthropic notes that Claude in Amazon Bedrock runs with zero operator access: Anthropic personnel have no access to the inference infrastructure. Logs go to CloudWatch and CloudTrail1. On Google Cloud, data handling is governed by Google Cloud, and its request-response logging service can log prompts and completions without giving Google or Anthropic access to the content2. Anthropic recommends keeping activity logs for at least 30 days on both12.
How to choose
- You need Batches, Files, server tools, the MCP connector or Managed Agents → the Claude API (or check Claude Platform on AWS).
- Claude must run inside AWS or Google Cloud, under that cloud's IAM, logging and agreements → Bedrock or Google Cloud, after checking the feature list against your code.
- You need data residency → compare Anthropic's US-only inference with each cloud's regional options and their 10% premiums.
- You have committed cloud spend → usage on that cloud counts toward it; usage on a direct Anthropic contract does not.
- You need more capacity or a failover path → use two platforms for the same model.
Two platforms behind one interface
Different model ids, request shapes and authentication on each platform are the cost of running two. A gateway that maps one model name onto each platform's id removes most of it. With Claude, failover between platforms for the same model is also the only failover that keeps Claude-specific prompts, tool use and coding agents behaving the same.
Connecting Claude to INFRO
INFRO connects Claude through your own Anthropic account and through Claude on Google Cloud. Your application names anthropic/claude-opus-5.5, and a route can order or weight the two connections with failure rules, so a 429 or an overload on one moves the request to the other. Claude Code and other Messages API clients reach both through INFRO's /v1/messages, sent to the provider as written, including thinking blocks and cache breakpoints. INFRO's Bedrock connection today speaks Bedrock's OpenAI-compatible Chat Completions endpoint and serves the OpenAI, xAI, Moonshot and MiniMax models Bedrock offers there — not Claude, which on Bedrock needs the Messages or Converse API — so Claude on Bedrock is not yet a connection INFRO offers.
Summary
All four routes serve the same Claude models. The Claude API has every feature first; Bedrock and Google Cloud put Claude inside your cloud with fewer features and a 10% premium for regional processing; Claude Platform on AWS combines Anthropic operation with AWS billing. Pick by the features your code uses and where your data must stay, and use a second platform for capacity and same-model failover.