Amazon Bedrock and Nova Explained: Amazon's Generative AI Stack for Beginners and Developers

Think of Bedrock as an "AI model marketplace plus toolkit" inside your AWS account:

• Pick a model from a catalogue (text, image, speech, embeddings).
• Send it a prompt through an API and pay per use, typically per input and output token.
• Add your own data safely, with tools for search over your documents, safety filters and agents.
• Keep data in AWS: AWS states that prompts and responses are not used to train the underlying models and are not shared with model providers.

For a business, this means you can add a chatbot, document summariser or AI agent to an app without managing infrastructure, and switch models later with little code change.

What is Amazon Nova?

Amazon Nova is Amazon's family of foundation models, first launched in December 2024 and refreshed as Nova 2 in December 2025. According to AWS's announcement and coverage of re:Invent 2025:

Reported Nova 2 capabilities include adjustable reasoning levels, built-in code interpreter and web grounding, support for remote MCP (Model Context Protocol) tools, and a one-million-token context window. AWS also introduced Nova Forge, a programme for organisations to build their own customised Nova variants; it is aimed at large enterprises, not hobbyists.

The first-generation Nova models (Micro, Lite, Pro, Premier, plus Canvas for images and Reel for video) remain in the Bedrock catalogue in many regions.

What are the main parts of the Bedrock stack?

AgentCore is AWS's platform for deploying and operating AI agents at scale (runtime, memory, identity, gateways to tools, observability). AWS has highlighted it heavily since 2025, but features change quickly, so check the AgentCore documentation for what is generally available in your region.

How do I start using Bedrock? (beginner steps)

1. Sign in to the AWS console and open Amazon Bedrock. Pick a region where the model you want is offered (US East (N. Virginia) and US West (Oregon) usually get models first; Asia Pacific (Mumbai) also offers Bedrock).
2. Check model access. AWS has simplified access so many serverless models are available by default, but some third-party models still require you to accept the provider's terms first.
3. Open the Chat / Text playground, choose a model such as Nova Lite, and type a prompt.
4. Set a budget in AWS Budgets before experimenting heavily.
5. Move to code with the AWS CLI or an SDK (Python boto3, JavaScript, Java and others) once you know which model fits.

Editor's tip: Start with the smallest, cheapest model that does the job. For classification, extraction or short summaries, a "Lite" or "Micro" class model is often enough, and costs a fraction of a frontier model.

For the technically curious: calling Nova through the Converse API

The Converse API gives the same request shape across models, so swapping models usually means changing only the model ID. First list the Amazon models available to you, then call one:

# List Amazon-provided models in your region
aws bedrock list-foundation-models --by-provider amazon \
  --query "modelSummaries[].modelId" --region us-east-1

# Send a prompt (replace MODEL_ID with an ID or inference profile from your console)
aws bedrock-runtime converse \
  --model-id MODEL_ID \
  --messages '[{"role":"user","content":[{"text":"Explain Amazon Bedrock in two sentences."}]}]' \
  --inference-config '{"maxTokens":300,"temperature":0.3}' \
  --region us-east-1

The same in Python with boto3:

import boto3

client = boto3.client("bedrock-runtime", region_name="us-east-1")

response = client.converse(
    modelId="MODEL_ID",  # e.g. a Nova Lite inference profile from your console
    messages=[{"role": "user", "content": [{"text": "Write a 3-line product description for a steel water bottle."}]}],
    inferenceConfig={"maxTokens": 300, "temperature": 0.5},
)

print(response["output"]["message"]["content"][0]["text"])
print(response["usage"])  # inputTokens / outputTokens, useful for cost tracking

Things developers commonly trip over:

• Inference profiles: many newer models must be called through a cross-region inference profile ID (often prefixed with a geography such as us. or global.) rather than the bare model ID. The console's model catalogue shows the right identifier.
• IAM permissions: your role needs bedrock:InvokeModel (and related actions for streaming or Converse) on the model or profile ARN.
• Streaming: use converse-stream / converse_stream for chat UIs so users see tokens as they arrive.
• Tool use: the Converse API supports passing tool definitions; the model returns a tool call, your code executes it, and you send the result back.

How RAG works with Knowledge Bases

1. You point a Knowledge Base at a data source such as an S3 bucket of PDFs.
2. Bedrock splits documents into chunks, converts them to embeddings with an embedding model, and stores them in a vector store.
3. At query time, the user's question is embedded, the closest chunks are retrieved, and they are passed to the model with the question.
4. The answer comes back with citations to the source chunks.

This is usually the fastest route to a "chat with our documents" assistant, and keeps answers grounded in your own data.

How much does Bedrock cost?

Bedrock pricing varies by model and region, so there is no single number. The main pricing modes are:

Amazon positions Nova models as low-cost options compared with many third-party frontier models. Check the Bedrock pricing page for the exact per-token rates in your region before you build cost estimates. New AWS accounts can use Free Tier credits for Bedrock.

Bedrock vs calling an AI provider directly

For teams already on AWS, Bedrock is usually the easier path for security reviews and procurement. For hobby projects that need one specific model's newest feature on day one, the provider's own API may be quicker.

What can you build with Bedrock and Nova?

Some practical, common use cases, with the Bedrock features they rely on:

Responsible AI basics before you ship

• Use Guardrails to block harmful content and redact personal data such as phone numbers or Aadhaar-like identifiers in prompts and outputs.
• Ground answers in your own documents with Knowledge Bases and show citations, to reduce made-up answers.
• Keep a human in the loop for anything customer-facing that carries legal, medical or financial weight.
• Log and monitor with CloudTrail and CloudWatch so you can audit who called which model and when.
• Evaluate before switching models: Bedrock offers model evaluation tools to compare outputs on your own test set.

FAQ

Is Amazon Nova free to use?

No, Nova models are billed per use in Bedrock. New AWS accounts can apply Free Tier credits to Bedrock usage, and trying Bedrock is one of the onboarding activities that earns extra credit.

What is the difference between Bedrock and Nova?

Bedrock is the platform (API, tools, security). Nova is a family of models made by Amazon that you can run on Bedrock alongside models from other companies.

Which Nova 2 models are available?

AWS announced Nova 2 Lite as available in Bedrock and Nova 2 Pro in preview in December 2025, along with Nova 2 Sonic for voice and Nova 2 Omni for multimodal tasks. Check the Bedrock console for current availability in your region.

Is Amazon Bedrock available in India?

Bedrock is offered in the Asia Pacific (Mumbai) region, but individual models vary by region. Some models are available to Indian users through cross-region inference.

Does AWS train models on my Bedrock data?

AWS states that Bedrock does not use your prompts and outputs to train the base models and does not share them with third-party model providers.

What is Bedrock AgentCore?

AgentCore is AWS's set of services for deploying and running AI agents securely at scale, covering runtime, memory, identity, tool gateways and observability.

Do I need machine learning experience to use Bedrock?

No. You can use the playground with no code, and the API calls look like any other web API. ML experience helps for fine-tuning and evaluation.

Sources

• https://aws.amazon.com/about-aws/whats-new/2025/12/nova-2-foundation-models-amazon-bedrock
• https://aws.amazon.com/bedrock/
• https://aws.amazon.com/bedrock/pricing/
• https://aws.amazon.com/ai/generative-ai/nova/
• https://charonhub.deeplearning.ai/nova-2-family-boosts-cost-effective-performance-adds-new-agentic-features/
• https://itdaily.com/news/cloud/aws-announces-new-ai-innovations/
• https://caylent.com/blog/aws-reinvent-2025-every-ai-announcement-including-amazon-nova-2-and-kiro

Featured image: Amazon. Source