Amazon Bedrock pricing in 2026: every model, hidden costs, and why your bill is higher than the pricing page suggests

The AWS Bedrock pricing page shows clean numbers. $3 per million input tokens for Claude Sonnet 4.6. $0.035 for Amazon Nova Micro. Multiply volume by rate, and you have an estimate. The estimate rarely matches the invoice. Across $15 billion in cloud and AI spend under management, CloudZero consistently sees organizations spending 1.5x to 2x their initial Amazon Bedrock cost estimates.

A team that budgets $50,000 annually ends up at $85,000. A team that budgets $200,000 ends up at $340,000. The overruns are not hidden fees. They are costs that the pricing page does not surface until they appear on the invoice: OpenSearch Serverless minimums for Knowledge Bases, embedding model inference for retrieval, Guardrails per-assessment charges, and the token amplification that happens when a Bedrock Agent chains multiple model calls to answer a single query.

This guide covers every Bedrock pricing tier, every model’s per-token rate, the costs that appear on the bill but not on the pricing page, how Bedrock compares to calling model providers directly, and the cost optimization tactics that reduce the gap between the estimate and the invoice.

For how Amazon Bedrock costs connect to broader AI spend and AI ROI, see CloudZero’s framework.

What is Amazon Bedrock?

For anyone asking what is Amazon Bedrock or what is AWS Bedrock: it is AWS’s fully managed AI service that provides API access to foundation models from Anthropic, Meta, Mistral, Cohere, AI21, Stability AI, Amazon’s own Nova family, and other providers through a single AWS endpoint. (In search, the service appears under variants like Bedrock AWS, Bedrock Amazon, and simply “Bedrock.”)

As of mid 2026, Amazon Bedrock models span 62 LLMs, 3 embedding models, and 6 specialized models across 42 AWS regions from 13 providers. The full AWS Bedrock models catalog includes every major model family: Claude (Anthropic), Llama (Meta), Mistral, Nova (Amazon), and Cohere Command.

Amazon Bedrock features include on-demand and batch inference, Amazon Bedrock Agents for multi-step task orchestration, Amazon Bedrock Knowledge Bases for retrieval-augmented generation, Amazon Bedrock Guardrails for content safety, model evaluation and fine-tuning, and cross-region inference. For an Amazon Bedrock overview of how all these components fit together, see the AWS documentation.

The value proposition: one API, one billing relationship (your AWS account), access to every major model family.

The cost reality: that single billing relationship includes on-demand inference, batch processing, provisioned throughput, Knowledge Bases, Agents, Guardrails, fine-tuning, model evaluation, and cross-region inference, each with its own pricing model.

Most teams start with on-demand token pricing and discover the other seven billing categories during a monthly bill review rather than at architectural planning.

For a broader view of what cloud computing is and how AWS Bedrock fits into the cloud service provider landscape, see CloudZero’s guides.

What are Amazon Bedrock’s 5 billing modes?

1. On-demand pricing (pay per token, no commitment)

The default and most common mode. You pay per input token and per output token with no upfront commitment. Rates vary by model. Here is the curated 2026 pricing table for the most-used models on Bedrock (us-east-1 region):

Prices are for us-east-1; other regions may vary

The 285x price spread between Nova Micro ($0.035 input) and Claude Fable 5 ($10.00 input) is the most important number in the Bedrock model pricing table. That spread defines the Bedrock cost per token range your team will operate within. Model routing (sending simple tasks to cheap models and complex tasks to expensive ones) is the single highest-impact Amazon Bedrock cost optimization lever. More on this below.

For AWS Bedrock model pricing across the full Claude family (AWS Bedrock Claude pricing specifically), see CloudZero’s Claude API pricing guide.

For a Bedrock pricing calculator approach, multiply your expected monthly token volume by the per-token rate for your chosen model, then add 50-100% for the hidden costs covered below.

2. Batch pricing (50% off, async processing)

Batch mode processes prompts asynchronously at 50% of on-demand rates. Claude Sonnet 4.6 drops from $3/$15 to $1.50/$7.50. Nova Pro drops from $0.80/$3.20 to $0.40/$1.60. For workloads that do not need real-time responses (data extraction, document analysis, bulk classification), batch is the simplest cost reduction.

3. Bedrock provisioned throughput (pay per hour, guaranteed capacity)

Reserved dedicated capacity billed per hour, not per token. Bedrock provisioned throughput pricing works best for high-volume, consistent workloads where you need guaranteed latency. Commitments of 1 month or 6 months; longer commitments reduce hourly rates. Break-even vs on-demand depends on volume: at fewer than 5 million requests/month, on-demand is typically cheaper.

Note on the Amazon Bedrock free tier: Bedrock does not have a permanent free tier. New AWS accounts created after July 15, 2025 receive up to $200 in credits ($100 on sign-up plus $100 for guided activities) usable across 200+ services including Bedrock. Credits expire after 6 months.

4. Prompt caching (up to 90% off input tokens)

Bedrock supports prompt caching for select models (Claude family, Nova family). Cached input tokens cost approximately 10% of the base rate. For applications with long, repeating system prompts (agentic workflows, Claude Code sessions, RAG pipelines), caching is the second-highest-impact cost lever after model routing. CloudZero’s Claude Code Plugin tracks caching effectiveness in real time.

5. Fine-tuning (per-token training + storage)

Custom model training charges per token of training data plus per-hour training time. Fine-tuned model storage adds monthly charges. Relevant for teams building domain-specific models but niche compared to on-demand and batch.

Those five billing modes are what the pricing page shows. What it does not show is where most of the budget overruns actually come from. 

What costs does the Amazon Bedrock pricing page not show?

These are the costs that drive the gap between the estimate and the invoice.

1. Knowledge Bases: the $345/month minimum

Amazon Bedrock Knowledge Bases connect your data to foundation models for retrieval-augmented generation (RAG). The default vector store for any AWS Bedrock Knowledge Base is Amazon OpenSearch Serverless, which requires a minimum of 2 OpenSearch Compute Units (OCUs) at $0.24/OCU/hour. That works out to $345/month even with zero query traffic. The meter is always running.

The 2026 alternative: Amazon S3 Vectors, generally available since December 2025, can cut vector storage costs by up to 90% versus specialized vector databases. For new Knowledge Bases, S3 Vectors should be the default unless you have a specific OpenSearch dependency. Amazon Bedrock Knowledge Bases pricing also includes embedding model inference (input tokens only) and Amazon Bedrock Data Automation for document parsing at $0.010 per page.

2. Amazon Bedrock Agents: token amplification

AWS Bedrock Agents orchestrate multi-step tasks by chaining model calls. A single user query can trigger 3-5 model invocations as the agent reasons, retrieves data, and generates a response. Each invocation bills separately. If the base model costs $3/$15 per MTok and the agent averages 4 calls per query, the effective cost per query is 4x what the pricing page suggests.

Gartner forecasts that supply chain management software with agentic AI alone will grow to $53 billion in spend by 2030. Agent token amplification is the cost challenge finance teams have not yet budgeted for. For how agentic workflows affect AI spend, see CloudZero’s guide. For teams running Claude Code AWS Bedrock sessions, agent-style chaining is the default pattern and caching becomes critical.

3. Amazon Bedrock Guardrails: per-assessment charges

AWS Bedrock Guardrails (content filtering, PII detection, prompt injection prevention) charge per text assessment. For high-volume applications running Guardrails on every request, the assessment cost can exceed the model inference cost. Budget for it separately.

4. Cross-region inference: no surcharge, but regional pricing applies

Cross-region inference routes requests across AWS regions at no additional routing fee. The source region’s pricing applies. Claude models with 1M token context windows (such as Fable 5) bill inputs beyond 200K tokens at approximately 2x the standard rate.US-only inference adds a 1.1x multiplier on all pricing categories. Unless compliance requires it, global routing saves 10%.

Those hidden costs raise a natural question: if Bedrock adds all this operational overhead, would it be cheaper to call model providers directly?

Amazon Bedrock vs direct API pricing: is the AWS markup worth it?

A common question: does Bedrock add a markup over calling providers directly?

Per-token rates match. The real cost difference is in the adjacent services: Knowledge Bases, Agents, Guardrails, and the AWS billing infrastructure add operational costs that calling the Anthropic API or OpenAI API directly does not.

The tradeoff: Bedrock gives you unified billing, VPC integration, IAM access controls, and AWS marketplace procurement. Direct API gives you lower operational cost but requires managing multiple billing relationships.

For teams running Claude on AWS specifically, CloudZero’s guide compares Bedrock, Claude Platform on AWS, and direct API costs side by side.

For how ChatGPT pricing and OpenAI API pricing compare to Bedrock rates, see CloudZero’s LLM API pricing comparison.

The direct API comparison answers the provider question. The next question is the AWS question: Bedrock or SageMaker?

AWS Bedrock vs SageMaker: when should you use which?

The AWS decision guide recommends Amazon Bedrock for teams that want to use existing foundation models without managing infrastructure, and Amazon SageMaker for teams training custom models or needing fine-grained control over the inference stack. Amazon Bedrock vs SageMaker is not an either/or decision. Most enterprise teams use both: Bedrock for prototyping and multi-model access, SageMaker for production ML pipelines.

Now that the pricing landscape is clear (5 billing modes, 4 hidden cost categories, the Bedrock vs direct API tradeoff, and the Bedrock vs SageMaker decision), here is how to reduce the actual bill.

How do you reduce AWS Bedrock costs? 7 ways

Here are practical tips to reduce Bedrock costs:

CloudZero’s CostFormation dimensions map every token to the workload that consumed it. CloudZero’s anomaly detection catches agent token amplification spikes and runaway Knowledge Base costs before the monthly bill arrives.

Organizations like Toyota, Duolingo, Coinbase, Shutterstock, Klaviyo, and Upstart manage AI spend through CloudZero at this level of granularity. CloudZero also tracks Bedrock spend alongside Google Vertex AI, GPU clusters, and 30+ other providers in a single view. It integrates with cloud cost management tools, cloud management software, and application monitoring tools most organizations already use. For teams running multi-cloud environments across AWS alternatives and other cloud service providers, CloudZero normalizes costs across all of them. and ask to see your Bedrock spend broken down by the dimensions that actually matter for AI ROI. You can also get a free cloud assessment today to see where your cloud and AI spend currently stand.

Frequently Asked Questions about Amazon Bedrock Pricing

Author Spotlight

Lyne Carolyne

Lyne Carolyne has several years of experience in AI and cloud economics and brings that understanding into the content she creates. Outside work, she's an avid explorer.