Enterprise deployment evaluation
Anthropic is not a processor of your data and needs no agreement with you. Your cloud provider’s own terms govern retention — AWS or Google Cloud. Both offer US region selection and a private network path.
No code execution, Files API, programmatic tool calling or batch.
Web search. Vertex AI offers a provider-side web search; Bedrock has none, so a separate search backend has to be configured. It is the only real feature gap between them.
Migration effort. Bedrock offers an Anthropic-native endpoint alongside its standard one, which makes moving code to or from the Claude Developer Platform less work.
Identity. One small point in Foundry’s favour against both: if your organisation already runs on Microsoft identity, federating sign-in is less work there.
The Azure-hosted option keeps prompts and responses within Azure and offers a US data zone. But Anthropic is a processor of your data under its own terms on either hosting option, which is what separates Foundry from the cloud providers.
Anthropic-hosted restores code execution, the Files API and programmatic tool calling — and sends everything to Anthropic. Azure-hosted has none of them. Batch is unavailable on both.
On Amazon Bedrock and Vertex AI, Anthropic supplies the model to your cloud provider and is not a processor of your data. Your agreement is with the cloud provider alone, and their terms decide what happens to what you send. On Foundry, Microsoft's own documentation states that Anthropic acts as an independent data processor for prompts and outputs, under Anthropic's own terms, on both hosting options.
Azure-hosted keeps the data in Azure: ingress, inference and storage at rest all run on Microsoft infrastructure in the Azure geography you pick, and a US data zone is available. The legal relationship does not follow the hardware: Anthropic, not Microsoft, is the processor, and Anthropic's terms govern retention.
The tools that run on Anthropic's servers are not available here.
Anthropic-hosted passes the request through to Anthropic's own service. The full tool set comes back and nothing stays in Azure. It is the same choice as going direct, made from inside Foundry.
Anthropic receives and processes all of it, under its own commercial terms. Zero Data Retention is available on the API, minus the features that have to keep data; the seat plans are not eligible for it at all.
Every tool in the table works. The seat plans add mobile and web access, sharing between users, audit logs and the Compliance API — the last two Enterprise-only, and none of them available on any managed-desktop option.
Remote code execution. Claude writes and runs code in a sandbox on Anthropic's servers to do the work — reading a spreadsheet, running the numbers, producing a chart. Without it Claude can still write the code, but someone has to run it.
Files API. Upload a document once and refer to it across many requests, with the file held on Anthropic's side. Without it, the content travels with every request that needs it.
Programmatic tool calling. Claude writes code that calls your own tools from inside that sandbox, instead of a round trip to the model for each call. It depends on code execution, so it is available wherever that is.
Batch processing. Queue a large volume of requests and collect the results later, usually at lower cost. Without it, high-volume back-office work runs as ordinary requests.
All four are services Anthropic operates, not abilities of the model. Claude itself is the same on every option; what differs is what runs alongside it.
| Capability | Claude Team and Enterprise | Claude Developer Platform | Amazon Bedrock | Google Cloud Vertex AI | Microsoft Foundry |
|---|