Amazon Bedrock
Listed at https://aws.amazon.com/bedrock/
💰 Token Pricing
| Type | Price | Note |
|---|---|---|
| Input | Claude Sonnet 4.5: $6/M, DeepSeek V3.2: $0.62/M, Mistral Large 3: $0.50/M, GPT-5.5: $5.50/M, GPT-5.4: $2.75/M, Grok 4.3: $1.25/M, Qwen3 235B: $0.23/M, Kimi K2.5: $0.60/M | per million tokens |
| Output | Claude Sonnet 4.5: $30/M, DeepSeek V3.2: $1.85/M, Mistral Large 3: $1.50/M, GPT-5.5: $33/M, GPT-5.4: $16.50/M, Grok 4.3: $2.50/M, Qwen3 235B: $0.91/M, Kimi K2.5: $3/M | per million tokens |
🤖 Supported Models (100)
✨ Pros
- ✓100+ models from 20+ providers: Anthropic Claude, Meta Llama, Mistral, DeepSeek, OpenAI GPT, Amazon Nova/Titan, Cohere, and more
- ✓Enterprise-grade security: IAM fine-grained permissions, VPC private networking, KMS encryption, HIPAA/SOC 2 certified
- ✓Deep AWS ecosystem integration: seamless with Lambda, S3, CloudWatch, SageMaker, IAM
- ✓Flexible pricing tiers: On-Demand pay-as-you-go, Batch at 50% discount, Provisioned Throughput, commitment discounts
- ✓AgentCore platform: build, deploy, monitor AI agents with any framework and model
⚠️ Cons
- ×Complex pricing varies by model, region, service tier, and commitment period
- ×Not directly accessible in mainland China: AWS China regions don't offer Bedrock yet
- ×Higher onboarding overhead: requires AWS account, IAM setup, service access requests
🎯 Best For
Enterprise customers on AWS; production apps needing multi-model flexibility; compliance-heavy workloads
💰 Pricing & Plans
| Model | Input ($/M tokens) | Output ($/M tokens) | Best For |
|---|---|---|---|
| Claude Opus 4.8 | $15.00 | $75.00 | Top-tier reasoning, agentic coding, long-form analysis |
| Claude Sonnet 4.6 | $6.00 | $30.00 | Production code generation, tool-use agents |
| Claude Haiku 4.5 | $1.00 | $5.00 | Latency-sensitive chat, high-volume lightweight tasks |
| DeepSeek V3.2 | $0.62 | $1.85 | Open-weights-class quality at frontier-budget price |
| Mistral Large 3 | $0.50 | $1.50 | European multilingual, code generation |
| GPT-5.5 | $5.50 | $33.00 | OpenAI flagship via Bedrock, multimodal native |
| GPT-5.4 | $2.75 | $16.50 | Cheaper GPT-5 family entry on Bedrock |
| Llama 3.3 70B Instruct | $0.72 | $2.28 | Open-weights meta-Llama on managed Bedrock |
| Amazon Nova Pro / Lite | $0.80 / $0.06 | $3.20 / $0.24 | AWS-native multimodal, cost-optimised |
🔧 API & Developer Experience
- •Unified Converse API: A single Bedrock Converse / ConverseStream endpoint abstracts model selection — the same request body and response schema drives Claude, Llama, Mistral, DeepSeek, Amazon Nova, Cohere, AI21, OpenAI GPT, and stability.ai Stable Diffusion. Switching models is a runtime arg, not a code rewrite, which makes A/B eval, traffic shifting, and per-prompt model routing tractable.
- •AWS SDK Integration: Bedrock is a managed AWS service — invoke through boto3 (Python), AWS SDK v2 (Java/Go/Node/Rust/PHP/.NET), AWS CLI, or Terraform/CloudFormation/CDK. Auth inherits from standard IAM SigV4, so roles, instance profiles, and cross-account assume-role policies work without separate API keys.
- •Model Customisation: Continued pre-training and fine-tuning are first-party — upload a labelled S3 dataset, kick off a Custom Model job, then serve it through a Provisioned Throughput endpoint. Provisioned Throughput can also import base model weights from Hugging Face via SageMaker JumpStart, or bring adapter weights trained elsewhere.
- •Agents & Tool Use: Native tool-use blocks in the Converse API, plus a Bedrock Agents service that orchestrates multi-step plans with action groups (Lambda-backed) and a Knowledge Base (RAG over S3 + OpenSearch / Pinecone / Redis Enterprise) — agents can also be wired into a flow editor for visual branching.
- •Guardrails & Safety: Bedrock Guardrails apply uniformly across models: content filters (hate, violence, sexual, misconduct, prompt-injection), denied-topic lists, sensitive-info redaction (PII, regex patterns), and grounding checks against the retrieved context for RAG pipelines. Same policy file applies to any underlying FM.
- •Streaming, Batch, Async Invoke: ConverseStream for SSE token-by-token responses; Batch mode (50% discount, results returned within 24h); InvokeModel async for offline bulk. Provisioned Throughput (hourly commit) is required for heavy-traffic baselines and custom models.
🛡️ Enterprise & Compliance Posture
Bedrock's value proposition versus raw model-API access is the AWS-grade compliance envelope wrapped around the underlying model. Every request sits inside IAM (fine-grained action-level permissions, service control policies, deny rules), VPC (private networking via VPC endpoints, no public egress), KMS (customer-managed encryption keys), and CloudTrail (per-call audit logs). HIPAA, SOC 1/2/3, ISO 27001/27017/27018, PCI DSS, FedRAMP Moderate, C5, K-ISMS, and MTCS Tier 3 certifications apply to the Bedrock service across most major regions; AWS Artifact makes the audit reports downloadable on demand. Data is not used to train the underlying models by default — Anthropic, Meta, Mistral and others each honour the AWS-side zero-retention contract for API payloads. Predictable latency and SLAs (99.9% monthly uptime for Bedrock InvokeModel in major regions, with Service Health Dashboard integration and enterprise support plans) round out the package for regulated workloads where consumer API tiers are explicitly out of bounds.
🌐 China Access & Latency
Amazon Bedrock is not currently available in the AWS China (Beijing) or AWS China (Ningxia) regions operated by Sinnet / Ningxia Western Cloud under a separate AWS account — mainland Chinese customers cannot open a Bedrock console, request model access, or invoke the Converse API from inside those regions. Production routing from mainland China to Bedrock's global endpoints (us-east-1, us-west-2, eu-west-1, ap-northeast-1, etc.) typically requires a stable overseas egress — Hong Kong, Tokyo, or Singapore — with observed round-trip latency of 280-500 ms depending on carrier and 6to4 / AnyCast path. Pricing and quotas are unchanged (you pay USD, no Alipay / WeChat Pay top-up, no domestic invoicing). For workloads with China-hosted users, the practical paths are: (1) run the API calls from an AWS account in Hong Kong or Singapore, fronting Bedrock in the same or a US region over AWS backbone (avoids the public internet entirely for sub-100 ms cross-region latency); (2) run Bedrock from a non-China AWS account entirely, accepting the corporate overseas egress; (3) substitute Alibaba Bailian (Tongyi / Qwen), Tencent hunyuan, or DeepSeek on Volcano Ark / siliconflow for workloads that must stay inside the Great Firewall. There is no official mirror in mainland China and no announced roadmap to launch Bedrock on Sinnet / Western Cloud.