Amazon Bedrock is a managed service that allows you to use large language models (LLMs) from various companies with a single AWS API and authorization system. Instead of paying API usage fees directly to Anthropic or Mistral, you go through AWS, which naturally raises the question: “Isn’t it more expensive due to the intermediary fee?”
The conclusion is not simple. In standard or global endpoints in US regions, many models have the same token pricing as direct API usage from the provider. On the other hand, if you choose endpoints that process data only within specific countries/regions, priority processing tiers, and Bedrock’s additional features, the actual billed amount can increase. Ultimately, the comparison is not just a single line of model price, but the sum of the model version, region, processing method, and surrounding AWS costs.
The prices in this article were compiled by checking the official price list published on September 2, 2026. Prices and model availability change frequently, so it is necessary to reconfirm the linked price list just before actual adoption.
Conditions to Match Before Comparison
All prices are compared in US dollars per 1 million input tokens and 1 million output tokens. Input tokens are prompts, documents, and conversation history sent to the model, and output tokens are the responses generated by the model.
Above all, products with only similar model names should not be treated as the same model. A one-to-one comparison is only possible when the identical version and service tier are confirmed. For open-weight models like Meta Llama, where the creator does not directly provide a general-purpose paid API, or models like Amazon Nova, which are primarily offered through Bedrock, there is no basis for comparison with a direct API.
Claude: Global prices are the same, regional fixed prices can be higher
According to Anthropic’s official price list, Claude Sonnet 4.5 is $3 for input and $15 for output, while Claude Haiku 4.5 is $1 for input and $5 for output. Bedrock’s global endpoints can also be configured to apply the same base pricing.
The difference appears when restricting data processing locations. Anthropic documentation explains that for Claude 4.5 and later models, Bedrock’s regional endpoints apply a 10% higher price than global endpoints. Therefore, if Sonnet 4.5 is fixed to a specific region, the simple converted price becomes approximately $3.30 for input and $16.50 for output.
| Model | Provider Direct API | Bedrock Global | Bedrock Regional Fixed | Interpretation |
| — | — | — | — | — |
| Claude Sonnet 4.5 | $3 / $15 | Approx. $3 / $15 | Approx. $3.30 / $16.50 | Global same, regional fixed approx. 10% increase |
| Claude Haiku 4.5 | $1 / $5 | Approx. $1 / $5 | Approx. $1.10 / $5.50 | Global same, regional fixed approx. 10% increase |
Here, the first number is the price per 1 million input tokens, and the second is for 1 million output tokens. Anthropic’s documentation as of September 2 states that the previously announced Claude Sonnet 5 price of $2 for input and $10 for output, which was set to expire on August 31, 2026, will be maintained as the standard price. However, some notices on the AWS pricing page still show the previous price increase warning, indicating a difference in document update times. If you are actually using Sonnet 5, it is safest to confirm the final price on the AWS console for the relevant endpoint.
Mistral: US prices are the same, but Tokyo prices differ
According to Mistral’s official price list, Mistral Large is $0.50 for input and $1.50 for output. Bedrock’s Mistral Large 3 also shows $0.50 for input and $1.50 for output in the US East, Ohio, and Oregon regions.
In the Tokyo region, it is $0.61 for input and $1.82 for output. This is approximately 21% higher than the US regions for the same request. This is not simply an “AWS intermediary fee” but rather a regional price difference published by AWS.
| Model | Mistral Direct API | Bedrock US | Bedrock Tokyo |
| — | — | — | — |
| Mistral Large 3 | $0.50 / $1.50 | $0.50 / $1.50 | $0.61 / $1.82 |
| Devstral 2 123B | Direct price reconfirmation needed | $0.40 / $2.00 | $0.48 / $2.40 |
Bedrock’s Mistral models also vary by service tier, such as Standard, Priority, and Flex, in addition to regions. The AWS price list also indicates model groups where Priority is 75% higher than Standard, and Flex is 50% lower than Standard. This is why real-time services, where low latency is crucial, and batch jobs, where processing can be slower, should not be calculated at the same unit price.
Why OpenAI, Llama, and Nova are difficult to put in the same table
Bedrock also offers OpenAI models, but you must first confirm whether the model name and deployment type exactly match OpenAI’s direct API products. For example, gpt-oss is an open-weight model, so it’s not a matter of comparing it to a product where OpenAI hosts the same model as a pay-as-you-go API under identical conditions. Bedrock pricing is the cost for AWS to host and operate the model, and it should be compared with the cost of your own GPUs or other hosting providers.
The same applies to Meta Llama. Since Meta does not directly sell a general-purpose paid API like Anthropic, you cannot calculate a “Bedrock premium rate compared to Meta’s direct price.” Amazon Nova is an AWS-native model, so there is no external direct API price. “Comparison impossible” does not mean it’s expensive, but rather that there is no one-to-one benchmark.
Calculating by Monthly Usage
Let’s assume a monthly usage of 10 million input tokens and 2 million output tokens.
| Model/Path | Calculation | Monthly Inference Cost |
| — | — | — |
| Claude Sonnet 4.5 Direct API | 10×$3 + 2×$15 | $60 |
| Claude Sonnet 4.5 Bedrock Global | 10×$3 + 2×$15 | Approx. $60 |
| Claude Sonnet 4.5 Bedrock Regional Fixed | 10×$3.30 + 2×$16.50 | Approx. $66 |
| Claude Haiku 4.5 Direct/Global | 10×$1 + 2×$5 | Approx. $20 |
| Mistral Large 3 Direct/Bedrock US | 10×$0.50 + 2×$1.50 | $8 |
| Mistral Large 3 Bedrock Tokyo | 10×$0.61 + 2×$1.82 | Approx. $9.74 |
This calculation only reflects standard inference tokens. It excludes prompt caching, long context, batch discounts, image/audio input, tool call costs.
Why AWS billing can be higher even if token prices are the same
The advantage of Bedrock is not just model invocation, but the ability to combine IAM permissions, CloudTrail auditing, Guardrails, Knowledge Bases, and Agents within an AWS account. However, these additional features are not free.
- Text and image policy checks for Bedrock Guardrails
- Embedding generation and vector search for Knowledge Bases
- Data layers like OpenSearch Serverless, Aurora, S3
- Lambda and external APIs invoked by Agents
- CloudWatch log and metric storage
- PrivateLink interface endpoints
- Inter-region or internet data transfer
- VAT and exchange rates
Therefore, a more accurate cost comparison formula would be:
Bedrock 총비용
= 모델 입력 토큰 비용
+ 모델 출력 토큰 비용
+ Guardrails·RAG·Agent 비용
+ 로그·스토리지·네트워크 비용Direct APIs also incur separate costs. If you build your own application gateway, secret management, audit logs, and vector database, you will need to account for those costs and operational personnel. Bedrock is an option to integrate some of these into AWS services, not a service that simply makes token prices cheaper.
When is a direct API advantageous?
If you only use one type of model and want to quickly apply the provider’s latest features, a direct API is simpler. It’s also easier to leverage the provider’s unique prompt caching, tools, batch processing, and discount agreements. If global processing is possible and your organization’s AWS integration requirements are low, there’s less reason to add an intermediary layer.
Conversely, if you need to switch between models from multiple providers within a single AWS account, or if IAM, CloudTrail, PrivateLink, and data regions are important, Bedrock’s value increases. Even if the token unit price is slightly higher, there can be operational benefits from reducing separate API keys, billing, and auditing systems.
Conclusion: AWS is not always an expensive intermediary
There are cases where Bedrock prices are the same as the provider’s direct API when matching the same model, version, and processing location. Key examples confirmed are Claude’s global processing and Mistral Large 3 in US regions.
However, if processing is fixed to a specific region, Claude’s price can increase by approximately 10%, and Mistral’s Tokyo price can be about 20% or more higher compared to US prices. Adding Guardrails, RAG, logs, and networking can further increase the overall AWS cost.
Therefore, the final decision should not be “AWS is an intermediary, so it’s always expensive,” but rather based on these three questions: Is it the same model and version? Is the data processing region the same? How many security and operational features beyond tokens are included? After matching these conditions, calculating with monthly input/output tokens will reveal the actual cost difference.
Leave a Reply