Amazon Bedrock API 2026: 100+ Models for the AWS Enterprise

What Is Amazon Bedrock?

Amazon Bedrock is not another AI model provider — it is a managed service on AWS that provides API access to foundation models from Anthropic, Meta, Mistral, DeepSeek, OpenAI (via AWS), Cohere, Stability AI, AI21 Labs, and Amazon's own Nova and Titan model families. It launched in 2023 and has since become the most comprehensive model catalog available through any single cloud provider.

The key insight: instead of signing up for 10 different API providers, managing 10 billing systems, 10 API keys, and 10 SDKs, Bedrock customers use their existing AWS account — the same IAM roles, VPC security groups, CloudWatch monitoring, and billing console they already use for EC2, Lambda, and S3.

By mid-2026, Bedrock has expanded to include 100+ models spanning:

• Anthropic: Claude Opus 4.8/4.7, Sonnet 4.6/4.5, Haiku 4.5, Mythos 5, Fable 5
• Meta: Llama 3.3 (70B/405B), Llama 4 (Scout/Sentry/Audience)
• DeepSeek: V3.2, V3.1, R1
• OpenAI: GPT-5.5, GPT-5.4
• Mistral: Large 3, Small
• Amazon: Nova Micro/Lite/Sonic/Pro/Premier, Canvas (image), Reel (video), Titan Text/Embeddings/Image Generator
• Cohere: Command R/R+, Embed v4
• Qwen3: 235B, 32B, Coder
• Others: Grok 4.3, Gemma 4, GLM 5, Jamba 1.5, Pixtral Large, Devstral 2, Kimi K2.5

This breadth makes Bedrock a compelling option for teams that need flexibility across model families without contract lock-in.

Pricing: Four Tiers of Flexibility

Bedrock's pricing is more complex than a pure-play API provider because it offers four consumption models plus regional variation. Here is the breakdown:

Tier	Description	Price Discount vs On-Demand	Best For
On-Demand	Pay per token, no commitment	Baseline (0%)	Evaluation, low-volume, variable workloads
Batch	Asynchronous bulk inference	50% off	High-volume non-real-time (classification, extraction)
Provisioned Throughput	Reserved model capacity (hourly)	Depends on commitment (1 month / 6 months)	Production workloads with steady throughput
Custom Model Import	Bring your own fine-tuned model	Compute-based pricing	Teams with proprietary fine-tuned models

Selected On-Demand prices for popular models (input / output per million tokens):

Model	Input Price	Output Price
Claude Sonnet 4.5	$6.00/M	$30.00/M
DeepSeek V3.2	$0.62/M	$1.85/M
Mistral Large 3	$0.50/M	$1.50/M
GPT-5.5	$5.50/M	$33.00/M
GPT-5.4	$2.75/M	$16.50/M
Qwen3 235B	$0.23/M	$0.91/M
Kimi K2.5	$0.60/M	$3.00/M
Grok 4.3	$1.25/M	$2.50/M

Prices are at parity with the model providers' own APIs, with no significant markup. Batch is the standout discount: at 50% off, it is the most cost-effective way to run large-scale non-real-time inference.

Enterprise-Grade Security and Compliance

This is Bedrock's strongest differentiator. Every other AI API provider in this comparison offers standard HTTPS encryption and basic authentication. Bedrock adds:

• IAM — fine-grained permissions per model per action (InvokeModel, ListFoundationModels)
• VPC — Bedrock can operate entirely within your VPC, data never traverses the public internet
• KMS — customer-managed encryption keys for all data at rest and in transit
• CloudTrail — every API call logged for audit trails
• CloudWatch — token usage, latency, error rates monitored with Alarms
• HIPAA Eligible — Bedrock is HIPAA eligible under AWS BAA
• SOC 2 / ISO 27001 — audited compliance
• Data Never Leaves AWS Network — inference traffic routes through AWS backbone, not the public internet

For financial services, healthcare, government, and regulated industries, this enterprise guardrail is the deciding factor between "we can use AI" and "we cannot."

Getting Started with Bedrock

Unlike pure-play API providers where you create an account and get an API key in 60 seconds, Bedrock requires an AWS account and service access enablement:

1. An AWS account with IAM user or role permissions (you must have existing AWS access)
2. Enable Bedrock in the AWS Console — request access to specific models (approval is usually instantaneous for most models; some require additional verification)
3. Configure IAM permissions — attach the AmazonBedrockFullAccess or a custom policy to your IAM role
4. Choose a service tier — Standard (default) or Flex (optimized for throughput)
5. Call the API via AWS SDK (Python boto3, AWS CLI, or REST)

The onboarding process is heavier than direct provider APIs, which is Bedrock's biggest trade-off: enterprise-grade security means enterprise-grade friction.

import boto3\nimport json\n\nbedrock = boto3.client(\n    service_name="bedrock-runtime",\n    region_name="us-east-1"\n)\n\nbody = json.dumps({\n    "anthropic_version": "bedrock-2025-01-01",\n    "max_tokens": 1024,\n    "messages": [{\n        "role": "user",\n        "content": "Explain Amazon Bedrock in one sentence."\n    }]\n})\n\nresponse = bedrock.invoke_model(\n    modelId="anthropic.claude-sonnet-4-5-v1",\n    body=body\n)\n\nresult = json.loads(response["body"].read())\nprint(result["content"][0]["text"])

API Surface: AWS SDK-Centric

Unlike OpenAI-compatible providers, Bedrock uses the AWS SDK (boto3 for Python, AWS CLI, or the Bedrock REST API). The request body differs by model family — Anthropic models use the Anthropic API format, Meta Llama uses a different schema, and Amazon Nova uses its own format. This means you cannot simply swap a base URL to switch from OpenAI to Bedrock — you must adapt request formatting per model family.

For teams already on AWS, this is natural. For teams coming from OpenAI SDK, there is a learning curve. Workarounds exist: some open-source libraries abstract the model-specific formatting (e.g., LiteLLM supports Bedrock as a provider, offering a unified interface).

Availability Around the World

Bedrock is available in 15+ AWS regions globally. However, AWS China regions are a notable gap:

• Beijing region (operated by Sinnet) — Bedrock not available
• Ningxia region (operated by NWCD) — Bedrock not available
• All other regions including us-east-1, eu-west-1, ap-northeast-1, ap-southeast-1 — Bedrock available

Chinese developers or enterprises in mainland China must access Bedrock through global AWS regions using a VPN or direct connect. For teams that need China-direct multi-model access, FreeModel at freemodel.dev/invite/FRE-7a3b6220 offers OpenAI-compatible endpoints with many of the same models and China-direct connectivity.

Pros and Cons

Pros

• ✅ 100+ models from 20+ providers — unmatched catalog breadth
• ✅ Enterprise security: IAM, VPC, KMS, CloudTrail, HIPAA eligible, SOC 2
• ✅ Deep AWS integration: Lambda, S3, CloudWatch, SageMaker, IAM
• ✅ Flexible pricing: On-Demand, Batch (50% off), Provisioned Throughput, Reserved
• ✅ Batch at 50% discount — most cost-effective bulk inference option in the market
• ✅ AgentCore for building and deploying AI agents at production scale
• ✅ Global infrastructure with regional endpoints for low latency

Cons

• ❌ Complex pricing: varies by model, region, tier (Standard/Flex/Priority), and commitment period
• ❌ Not available in China: requires VPN/proxy to access global regions
• ❌ Higher onboarding overhead: AWS account, IAM setup, model access requests
• ❌ No OpenAI-compatible endpoint: requires AWS SDK, model-specific request formatting
• ❌ No free tier: must pay On-Demand or commit to Provisioned Throughput
• ❌ No public affiliate program

Use Case Recommendations

Use Case	Recommended Approach	Why
Enterprise production on AWS	Bedrock On-Demand or Provisioned	Deep AWS integration, IAM/VPC, compliance
Bulk offline processing	Bedrock Batch (50% off)	Half price for async inference
Multi-model experimentation	Bedrock (single API, 100+ models)	Zero multi-provider management overhead
Individual / startup (non-AWS)	Direct provider API or aggregator	Lower friction, no AWS account needed
China-based development	FreeModel or direct China provider	Bedrock not available in China regions
OpenAI-compatible drop-in	Direct provider or FreeModel	Bedrock uses AWS SDK, not OpenAI standard

Comparison: Bedrock vs Direct APIs vs Aggregators

Dimension	Amazon Bedrock	Direct Provider API	Aggregator (e.g. FreeModel)
Models	100+	1-50	50+
API Format	AWS SDK (model-specific)	Provider-specific or OpenAI-compat	OpenAI-compatible
Auth	AWS IAM + Signature v4	API key	API key
Onboarding	Complex (AWS account, IAM, enable)	Simple (register + key)	Simple (register + key)
China Access	Not available	Varies	Direct
Enterprise Security	IAM/VPC/KMS/HIPAA/SOC2	Basic HTTPS	Basic HTTPS
Pricing	Parity with direct ($0.23-$33/M)	Direct rate	Slight markup or bundling
Batch Discount	50% off (Batch inference)	Rare	Rare
Affiliate Program	No	Some	FreeModel: Yes

FAQ

Q: What models are available on Amazon Bedrock?

A: Bedrock offers 100+ models from 20+ providers including Anthropic Claude (Opus 4.8, Sonnet 4.6, Haiku 4.5, Mythos 5, Fable 5), Meta Llama 3.3/4, Mistral Large 3, DeepSeek V3.2/R1, OpenAI GPT-5.5/5.4, Amazon Nova/Titan, Cohere Command R/Embed, Qwen3 235B, Kimi K2.5, Grok 4.3, Gemma 4, GLM 5, Jamba 1.5, and Pixtral Large.

Q: Can I use Amazon Bedrock from mainland China?

A: No. Bedrock is not available in AWS China regions. You need a global AWS region accessed via VPN or proxy. For a China-direct alternative bundling similar models, see FreeModel at freemodel.dev/invite/FRE-7a3b6220.

Q: Is there a free tier for Bedrock?

A: No permanent free tier. AWS Free Tier provides up to $200 in new customer credits (6 months), which Bedrock usage consumes. Once exhausted, On-Demand rates apply.

Q: Does Bedrock support batch inference?

A: Yes. Batch inference is Bedrock's async bulk processing mode at 50% off On-Demand prices. Output goes to an S3 bucket. Ideal for high-volume non-real-time workloads like document processing, classification, and content moderation.

Q: Is Bedrock more expensive than calling APIs directly?

A: No — Bedrock prices are at parity with direct provider APIs for most models. The convenience of a single API for 100+ models comes at zero price premium. Batch is actually cheaper (50% less) than direct APIs typically offer.

Q: Does Amazon Bedrock have an affiliate program?

A: No. AWS does not operate a public affiliate program for Bedrock. For teams that want a multi-provider aggregator with an affiliate program, consider FreeModel.

Conclusion

Amazon Bedrock occupies a unique position in the AI API landscape. It is not the cheapest option (though Batch comes close), not the simplest to set up (AWS onboarding is real friction), and not accessible from China. What it offers instead is something no pure-play provider can match: 100+ models from 20+ providers, enterprise-grade security with IAM/VPC/KMS, deep AWS ecosystem integration, and Batch pricing at 50% off.

For enterprises already on AWS — especially those in regulated industries — Bedrock is the obvious choice. For individual developers, startups without AWS infrastructure, or teams that need a simple OpenAI-compatible drop-in, direct provider APIs or an aggregator like FreeModel are more practical.

If you are evaluating multi-provider access and want to compare Bedrock against an OpenAI-compatible aggregator with China-direct support, FreeModel bundles Claude, DeepSeek, GPT, Mistral, and many more models behind a single API key with deterministic model selection and no AWS account needed.

Comparison Table (Final)

Provider	Models	Pricing	China Access	Enterprise	Best For
Amazon Bedrock	100+	On-Demand + Batch (-50%)	Not available	IAM/VPC/KMS/HIPAA	AWS enterprise
FreeModel	50+	Model-dependent	Direct	Basic	Multi-provider, China direct
OpenAI	8	$0.15-$60/M	Proxy required	Basic	Standard API devs
Anthropic	~10	$3-$50/M	Proxy required	Basic	Claude model users
Groq	~8	$0.01-$0.59/M	Proxy required	Basic	Ultra-fast inference