Amazon Bedrock
One API to call foundation models from several providers, open-weight models included. You pay per token, and three choices decide most of the bill and of the compliance story.
Your app talks to one API. Bedrock routes to the model.
Prompts and responses are not used to train the models.
No rewrite to move an app onto Bedrock.
Converse
client.converse( modelId="…", messages=[…])
One request shape for every model on Bedrock. Switching model is a change of ID.
Chat Completions, Responses
OpenAI(base_url= ".../openai/v1")
Point the OpenAI SDK at Bedrock and keep your code. Works with gpt-oss and the other models that support it.
Messages
POST .../anthropic /v1/messages
The native Claude request format, for apps and tools built on the Anthropic SDK.
Same IAM, logging and guardrails behind each API. Which API a model supports is listed per model in the docs.
Open models, without running GPUs.
Availability varies by region: check yours before you plan on a model.
Most of the cost and the compliance comes from here.
Which model, for which step
Price per token grows with model size. Small models for simple steps, the large one only where it changes the result.
Where it runs
The inference profile decides which regions may serve a request. Agree it with your compliance team before the first model goes live.
How you pay
Batch costs half for jobs that can wait. Prompt caching cuts the price of a long prompt you resend; the ratio depends on the model.
Model sizes and the cached-input bar are illustrative shapes, not prices.
A Bedrock bill, in one line.
End-to-end agentic AI use cases
From business need to production, customer-facing included, with evaluation and ops to match.
Read the approach →The one-page Amazon Bedrock field guide.
Best practices, cost traps and performance checks, in the order you should run them. Printable, no fluff.