Skip to the page
Tecyfy

AI deployment interview question · medium

A client on AWS wants to use Claude. Should they call it through Amazon Bedrock or through Anthropic's own API? How do you decide?

The short answerThe model and, on Bedrock's global endpoint, the token price are the same, so decide on what surrounds the model.

The full answer

The model and, on Bedrock's global endpoint, the token price are the same, so decide on what surrounds the model. Choose Bedrock when the client's identity, logging, networking and procurement already run in AWS: access is IAM, calls land in CloudTrail, usage is on the AWS bill, and the model maker cannot see the prompts. Choose Anthropic's API when the product needs features Bedrock lacks, such as web search, code execution, the Files API or Message Batches, or wants them the day they ship. A third option, Claude Platform on AWS, is Anthropic's API billed through AWS Marketplace.

1. The questions to ask the client

QuestionPoints to
Must every vendor pass a security review, and is AWS already approved?Bedrock
Does the product use web search, code execution or files on the API side?Claude API
Must processing stay in the EU (or Japan, Australia)?Bedrock regional profile, +10%
Is there an AWS spending commitment to use up?Bedrock or Claude Platform on AWS
Is half-price batch processing a big part of the cost?Claude API's Message Batches (Bedrock has its own batch inference)
Does the team need new features on day one?Claude API or Claude Platform on AWS

2. The code barely differs

from anthropic import Anthropic, AnthropicBedrockMantle

direct = Anthropic()                                     # API key
bedrock = AnthropicBedrockMantle(aws_region="us-east-1") # AWS credentials

for client, model in [(direct, "claude-sonnet-5-5"),
                      (bedrock, "anthropic.claude-sonnet-5-5")]:
    client.messages.create(model=model, max_tokens=512,
                           messages=[{"role": "user", "content": "Hello"}])

So for plain messages and tool use the choice is reversible. Each API-only feature adopted makes it less so.

3. Things that surprise teams later

  • Quotas are per model and region on Bedrock, and output tokens of newer Claude models count several times against them. See Rate Limit.
  • Retirement dates are set per platform: the same model can retire on a different day on Bedrock. See Model Deprecation.
  • An agent built on server tools must be rebuilt with your own tools to move to Bedrock.

4. A recommendation

For an AWS-first enterprise, start on Bedrock and keep to messages and tool use; for a product team that needs the hosted tools, the Claude API. Write the decision down with the feature list it depends on, so the next review knows what would change it. See Amazon Bedrock.

What interviewers want: same model and price, a decision on governance, features and procurement, the specific features missing on Bedrock, regional processing options and their cost, the near-identical SDK code, platform-specific quotas and retirement dates, and the third option of Claude Platform on AWS.

Last updated 22 October 2026

Sources

Where each fact comes from

  1. 1Claude in Amazon Bedrock
  2. 2Data protection
  3. 3Data residency