AWS Bedrock Pricing: Every Model and Feature Cost

AWS Bedrock gives you access to foundation models from Anthropic, Meta, Mistral, Stability AI, Amazon, and others through a single API. The pricing model is straightforward — pay per token for text models, per image for image models — but the total cost depends heavily on which model you choose, how you structure prompts, and which Bedrock features you layer on top. A 10x price difference exists between the cheapest and most expensive text models for the same task.

TL;DR: Bedrock text models range from $0.0002/1K input tokens (Titan Express) to $0.015/1K (Claude Opus). Most production workloads should use Claude Sonnet ($0.003/1K input) or Llama 3.1 70B ($0.00099/1K). Batch inference gives 50% off, Provisioned Throughput saves 20-40% for sustained loads. Knowledge Bases, Agents, and Guardrails each add per-query fees on top of model costs.

---

Text Model Pricing

Anthropic Claude Models

Meta Llama Models

Mistral Models

Amazon Titan Models

---

Image Model Pricing

---

Embedding Model Pricing

Embeddings are extremely cheap — processing 1 million documents (1K tokens each) costs $20 with Titan Embeddings.

---

Pricing Modes

On-Demand (Default)

Pay per token/image with no commitment. Best for variable or low-volume workloads.

Batch Inference

Batch inference is ideal for document processing, content generation pipelines, and any workload that doesn't need real-time responses.

Provisioned Throughput

Provisioned Throughput guarantees a specific number of model units dedicated to your account. See the Bedrock user guide for configuration details. Best for sustained high-volume production workloads.

Prompt Caching

For applications with large, repeated system prompts, caching reduces input costs dramatically.

---

Feature Pricing

Knowledge Bases (RAG)

The vector store is often the largest Knowledge Base cost. OpenSearch Serverless minimum is $701/month.

Agents

Agents consume additional tokens for reasoning, tool selection, and orchestration — typically 2-5x the tokens of a direct model call.

Guardrails

Guardrails are applied in addition to model costs. For high-volume applications, guardrail costs can be significant.

Model Evaluation

---

Real-World Cost Examples

---

Cost Optimization Quick Wins

1. Use the smallest model that meets quality requirements — Haiku costs 95% less than Opus
2. Enable batch inference for non-real-time workloads — instant 50% savings
3. Implement prompt caching for repeated system prompts — up to 90% input savings
4. Set max_tokens appropriately — don't allocate 4,096 tokens for tasks that need 200
5. Route requests by complexity — simple tasks to Haiku, complex to Sonnet

---

Related Guides

• AWS Bedrock Cost Optimization Guide
• AWS Bedrock vs OpenAI Pricing
• AWS Bedrock LLM Models Guide
• AWS Bedrock Batch Inference Guide
• AWS Bedrock vs SageMaker

---

FAQ

Which Bedrock model offers the best price-to-performance ratio?

Claude Sonnet for general tasks, Llama 3.1 70B for cost-sensitive workloads, and Claude Haiku for high-volume simple tasks. Titan Text Express is cheapest overall but less capable.

Is Bedrock cheaper than calling model APIs directly?

Bedrock pricing matches or slightly exceeds direct API pricing (e.g., Anthropic API). The value is in unified billing, VPC integration, Guardrails, Knowledge Bases, and not managing API keys across multiple providers.

How do I estimate my monthly Bedrock costs?

Estimate: (average input tokens per request x input price) + (average output tokens per request x output price) x monthly request volume. Add 20-30% for agent orchestration overhead if using Agents. Add vector store costs if using Knowledge Bases.

---

Lower Your Bedrock Costs with Wring

Wring helps you access AWS credits and volume discounts to lower your Bedrock costs. Through group buying power, Wring negotiates better rates so you pay less per model inference.

Start saving on Bedrock →