The Amazon Bedrock console includes a playground: an interactive, in-browser environment where you pick a model, type a prompt, adjust settings, and read the response. For enterprises evaluating Claude, this is the natural first stop. Product managers and analysts can validate whether a use case works at all; engineers can converge on a system prompt and response format before committing anything to code. Because playground requests are ordinary model invocations under your AWS account, they exercise the same access controls and billing as production traffic — just at hand-typed volume.
What the playground is good for
System prompt iteration. Most of the effort in an enterprise Claude deployment goes into the instructions, not the plumbing. The playground lets you rewrite instructions and rerun the same user input repeatedly, at conversational speed, with no deploy cycle.
Model comparison. Bedrock exposes several Claude models — on the current surface: anthropic.claude-fable-5, anthropic.claude-opus-4-8, anthropic.claude-sonnet-5, and Haiku 4.5 among them. Running the same prompt against a cheaper and a more capable model is the quickest way to discover whether your task actually needs the expensive one. (See the platform overview for the model lineup.)
Response format validation. If your application needs JSON, a fixed schema, or a particular tone, verify in the playground that your instructions reliably produce it before you build parsing code around the assumption.
Prerequisites: model access and IAM
Two things must be true before the playground works. First, Claude model access must be enabled in the account — see requesting Claude model access; the first use of an Anthropic model triggers the Marketplace subscription and, on the classic surface, a one-time use-case form. Second, the console user needs IAM permission to invoke models. AWS's documented minimal playground policy allows bedrock:InvokeModel and bedrock:InvokeModelWithResponseStream on both foundation-model ARNs (arn:aws:bedrock:*::foundation-model/...) and inference-profile ARNs — the latter matters because invoking through a cross-region inference profile requires permissions on both resource types. Broad managed policies like AmazonBedrockFullAccess also work for experimentation, though production roles should be scoped down (see least-privilege IAM).
Know which knobs won't transfer to code
The playground exposes inference settings, but recent Claude models have changed what is adjustable: on Claude Sonnet 5 and Opus 4.7+, sampling parameters like temperature at non-default values are rejected by the API, and thinking behavior is controlled through adaptive thinking and effort rather than manual budgets. Treat the playground as a place to converge on prompt and model choice; confirm parameter behavior against the current API documentation when you translate to code.
From playground to Python
Once the prompt behaves, moving to code is mechanical. The current "Claude in Amazon Bedrock" surface uses the Messages API with the same request shape as Anthropic's first-party API:
from anthropic import AnthropicBedrockMantle
client = AnthropicBedrockMantle(aws_region="us-east-1")
msg = client.messages.create(
model="anthropic.claude-sonnet-5",
max_tokens=1024,
system="You are a claims-triage assistant. Reply in JSON.",
messages=[{"role": "user", "content": "Categorize: cracked windshield"}],
)
print(msg.content[0].text)Install with pip install -U "anthropic[bedrock]". The system field carries the system prompt you refined in the playground, and the model ID keeps its anthropic. prefix. Note that the ant CLI does not support Bedrock — use an SDK or cURL.
Where to go next
When you're ready to productionize, read your first API call on Bedrock and IAM service roles for compute calling Bedrock. The quickstart covers the same journey on all four platforms.