What Is Amazon Bedrock? — Bedrock Foundations
← Back to Dashboard
1. What Is Amazon Bedrock?2. Core Services and Request Flow
What Is Amazon Bedrock?
📚 Bedrock Foundations⏱ 8 min⭐ 60 XP⌨ Hands-on lab
Managed Foundation Models on AWS
Amazon Bedrock is AWS's fully managed foundation model platform. You call frontier models through one unified API without managing GPU fleets, model-serving containers, or patching inference infrastructure. Think of it as "RDS for foundation models" - AWS runs the heavy machinery, you consume capability through an endpoint that already speaks IAM, CloudWatch, CloudTrail, and KMS.
The Provider Catalog
One API surface, many model families. Availability depends on region and per-account model access, but the catalog spans:
| Provider | Model families | Typical strengths |
|---|---|---|
| Amazon | Nova 2 (Lite, Pro, 2.5 Sonic), Nova (Micro, Lite, Pro, Premier, Sonic, Canvas, Reel), Titan | Configurable reasoning effort, speech-to-speech, multimodal, cost-efficient text & embeddings |
| Anthropic | Claude 5.x (Fable 5.1, Opus 5.5, Sonnet 5.5, Haiku 5.5 — note Claude 3.5 Sonnet is retired EOL) | Frontier reasoning, agentic coding & tool loops, long context |
| Meta | Llama 4 (Scout 17B, Maverick 17B), Llama 3.3 70B | Open-weight MoE & dense families, long-context multimodal |
| DeepSeek | DeepSeek-R1 (us.deepseek.r1-v1:0), DeepSeek-V3.2 | Chain-of-thought reasoning & high-efficiency code/math |
| Mistral AI | Mistral Large, Pixtral Large, Ministral | Strong quality-to-cost ratio, multimodal, European provider |
| Cohere | Command, Embed, Rerank 3.5 | Enterprise RAG: multilingual embeddings and neural reranking |
| Google DeepMind | Gemma (Gemma 4 31B, 26B-A4B, E2B) | Reasoning, multimodal, open-weight, native function calling |
| xAI | Grok (Grok 4.7) | Configurable reasoning effort, agentic tool use, token efficiency |
| Others | AI21, Stability AI, Luma AI, OpenAI GPT-OSS, and more | Specialised text, image/video generation, open-weight reasoning |
Three Ways to Pay for Inference
| Mode | How it works | Use when |
|---|---|---|
| On-Demand | Pay per input/output token, no commitment | Default for interactive apps and spiky traffic |
| Batch inference | Asynchronous jobs against JSONL files in S3 at a 50% discount vs on-demand | Large offline workloads: classification, enrichment, evals |
| Provisioned Throughput | Reserved model capacity for a fixed term | Steady high-volume traffic and custom models |
Why Not Just Host a Model Yourself?
| Approach | You manage | Trade-off |
|---|---|---|
| Self-hosted on EC2/EKS | GPUs, serving stack, scaling, patching, model weights | Maximum control, maximum ops burden |
| SageMaker endpoints | Instance selection, containers, autoscaling policies | Great for custom/proprietary models you train |
| Bedrock | Only your application and prompts | Fastest path to production with AWS-native governance |
Key idea: Bedrock is not just text generation. It is a full platform - inference (Runtime APIs), safety (Guardrails), retrieval (Knowledge Bases), agents (AgentCore), evaluation, and governance - and every piece plugs into the AWS controls your organisation already audits.
⌨ HANDS-ON LABVerify Bedrock Access from CLI
⭐ +150 XPConfirm your AWS identity and Bedrock model visibility before building anything.
1Check your AWS CLI installation.
2Verify the active IAM identity.
3List foundation models in your target region.
OBJECTIVE 1 / 3 — type "hint" if stuck
🧪 Knowledge Check
Press 1-4 to select1 of 4
What is Bedrock's core value for engineering teams?
You train every model from scratch
Managed access to multiple FM providers through AWS APIs
A desktop-only AI IDE
Only image generation