πΊπΈ USA Β· AWS Bedrock
Status: π© COMPLETE π¦ LIVING Last updated: 2026-06-26 Plain-English tagline: AWSβs managed AI model catalog. One API, many models β Claude, Llama, Mistral, Cohere, Amazon Nova, all accessible via AWS infrastructure with AUS data residency.
Front-matter facts
| Field | Value |
|---|---|
| Vendor | Amazon Web Services (Seattle, USA) |
| Country / origin | πΊπΈ USA |
| Recommended for Australian users? | β
Yes β Bedrock available in Sydney (ap-southeast-2) for AUS data residency |
| Privacy summary | No training on customer data (contractually); tenant-isolated; AUS data residency via Sydney; IRAP compatible |
| Free tier | Limited free quota via AWS Free Tier |
| Paid tiers | Pay-per-token + provisioned-throughput (reserve capacity for guaranteed latency); AUD via AWS Australia billing |
| First released | April 2023 |
| Last reviewed | 2026-06-26 |
| Official site | https://aws.amazon.com/bedrock |
What it is
AWS Bedrock is Amazonβs managed AI model platform. Instead of using each AI vendorβs API separately, Bedrock provides one unified AWS-native API to call models from multiple providers:
- Anthropic Claude (Opus / Sonnet / Haiku / Fable) β Amazonβs largest AI partnership (US$8B+ investment)
- Meta Llama (4 / 5 family, including Llama Code, Llama Guard)
- Mistral (Large, Codestral, Small)
- Cohere (Command R / R+, Embed, Rerank)
- Amazon Nova (Micro / Lite / Pro / Premier / Canvas / Reel / Sonic)
- Stability AI (Stable Diffusion, Stable Image)
- AI21 Labs (Jamba)
- Plus: Bedrock-hosted versions of various models
Why use Bedrock instead of vendors directly:
- AWS-native β same AWS account, same IAM, same billing
- AUS data residency via Sydney region
- No-train guarantees apply uniformly
- Switch models easily β change a parameter, not your whole integration
- IRAP compatible for AUS government workloads
- Bedrock Knowledge Bases β built-in RAG with Aurora / OpenSearch
- Bedrock Agents β orchestrate multi-step agent workflows
- Bedrock Guardrails β built-in content / safety controls
What youβd use it for
- Youβre an AWS shop and want AI without leaving AWS
- AUS data residency for AI (Sydney is the strongest AUS-cloud AI option)
- Multi-model strategy β easily switch between Claude / Llama / Nova
- RAG (Retrieval-Augmented Generation) β Bedrock Knowledge Bases handles the wiring
- Agentic workflows β Bedrock Agents with tool use
- Production AI for regulated AUS industries β banking, government, healthcare
- Enterprise compliance β IRAP, HIPAA, SOC 2, GDPR all wrapped together
When NOT to use Bedrock:
- For consumer chat (use Claude.ai / ChatGPT / Gemini direct)
- For tiny prototypes (direct Anthropic / OpenAI APIs simpler)
- When non-AWS infrastructure makes more sense
How to use from Australia
- AWS account in
ap-southeast-2(Sydney) for AUS data residency - Navigate to Amazon Bedrock in console
- Request access to models you want (Claude / Llama / Nova / etc.)
- Most are auto-approved for established accounts
- Some require justification form (1-2 days)
- Use Bedrock console playground OR call via Bedrock API
- Billed via standard AWS billing in AUD
Example API call (Python via boto3)
import boto3
client = boto3.client('bedrock-runtime', region_name='ap-southeast-2')
response = client.invoke_model(
modelId='anthropic.claude-sonnet-4-6-20260105-v1:0',
body=json.dumps({"messages": [{"role": "user", "content": "Hello"}]})
)What it costs
Pay-per-token (most use)
Examples (verify current β Bedrock pricing changes):
- Claude Sonnet 4.6 via Bedrock: ~US15 per million input/output tokens
- Claude Haiku 4.5 via Bedrock: ~US5 per million
- Llama 4 70B via Bedrock: ~US0.45 per million
- Amazon Nova Micro: ~US0.14 per million (extremely cheap)
- Mistral Large via Bedrock: ~US6 per million
- Cohere Command R+: ~US10 per million
Provisioned Throughput
- Reserve dedicated capacity for guaranteed latency
- Monthly commit; significant discount for steady high-volume workloads
Image / video models
- Per-image or per-second pricing
- Stable Diffusion XL: ~US$0.018-0.072 per image (varies by step count)
AWS Free Tier
- Limited free trial quota across most Bedrock models
How it compares to alternatives
| Aspect | AWS Bedrock | Azure OpenAI Service | Vertex AI | Direct vendor APIs |
|---|---|---|---|---|
| Multi-vendor model catalog | Best (broadest) | OpenAI + MAI + Llama + Mistral | Gemini + Llama + Claude + Mistral | One per vendor |
| Cloud-native integration | Best for AWS | Best for Azure | Best for GCP | None |
| AUS data residency | Yes (Sydney) | Yes (Australia East) | Yes (Sydney + Melbourne) | Limited |
| Pricing markup over direct | Some (10-20%) | Some | Some | Lowest |
| Enterprise terms / SLAs | AWS enterprise terms | Microsoft enterprise terms | Google enterprise terms | Per-vendor |
| Best for | AWS shops + multi-model + AUS data residency | Microsoft shops + GPT access | Google shops + Gemini access | Vendor-direct optimisation |
For AUS organisations needing AI with data residency, Bedrock in Sydney is the strongest pick.
Privacy / data handling
- No training on customer data β contractually committed across all Bedrock models
- Customer prompts / responses not retained beyond whatβs required to serve the request
- AUS data residency in Sydney region
- IRAP compatible for AUS government
- HIPAA, GDPR, SOC 2, ISO 27001 compliant
- Tenant-isolated by AWS account
- Bedrock Guardrails β additional content / safety controls per use case
- CloudTrail logging of all Bedrock API calls
Recent changes
- 2026: Amazon Nova Premier added; broader model catalog
- 2025: Bedrock Agents matured; Knowledge Bases improved
- 2024: Bedrock expanded broadly; AUS Sydney availability for most models
- April 2023: Bedrock launched
Gotchas
- Model availability varies by region β not all models are in
ap-southeast-2Sydney; some may be US-only. Verify your specific models. - Quota management is real β start with small quotas, request uplift for production
- Pricing markup vs direct vendor is modest but real β for cost-extreme optimisation, direct APIs might win
- Cross-region inference is supported but adds latency + egress costs
- For tiny prototypes, direct Anthropic / OpenAI APIs are simpler β Bedrock setup overhead matters for small projects
- Bedrock Knowledge Bases has its own infrastructure costs (Aurora / OpenSearch) on top of model costs
- AUS billing native via AWS Australia; verify tax-invoice handling for ABN-registered businesses
See also
- AWS overview π© π¦
- AWS SageMaker π₯ β for ML training (not inference)
- Amazon Q Business π© π¦
- Amazon Q Developer π© π¦
- Amazon Nova π© π¦
- Claude models π© π¦ β accessible via Bedrock
- Llama π© π¦ β accessible via Bedrock
- Mistral models π₯ β accessible via Bedrock
- Cohere π© π¦ β accessible via Bedrock
- Azure OpenAI Service π₯ β comparable Microsoft offering
- Vertex AI π₯ β comparable Google offering
- Claude API overview π© π¦
- OpenAI API π₯
- RAG β Retrieval Augmented Generation π©