Amazon Bedrock pricing is a price per 1,000 input and output tokens by model, modality and service tier, with batch at half the rate, cache reads at a quarter of the input price and provisioned throughput per model unit and hour, plus per token for fine-tuning and per month for stored custom models; SageMaker AI charges per instance-second by component, notebooks, training, processing, endpoints, and for Serverless Inference per GB-second, plus storage and data transfer. On the bill, Bedrock is dominated by output tokens, which cost several times the input per token, and SageMaker by the endpoints and notebooks that run around the clock whether requests come or not.
Pricing dimensions
The two services are two models of the same thing: Bedrock sells model invocations and measures in tokens, SageMaker sells compute time and measures in instance-seconds. Every item carries its own usage type; on Bedrock the model sits in the usage type, on SageMaker the component and the instance type.
| Dimension | Unit | What drives it | Optional |
|---|---|---|---|
| Bedrock on-demand | 1,000 input tokens and 1,000 output tokens per model, output dearer; image, video and audio per modality at own rates | Every invocation, prompt and context as input, answer as output | No |
| Bedrock service tiers | Standard; Flex at the lower rate with longer response time; Priority at the higher rate with precedence; Reserved | The choice per invocation or application | Yes |
| Bedrock batch | Half the rate per token for jobs that run asynchronously through S3 | Bulk processing without a response time | Yes |
| Bedrock prompt caching | Cache reads at a quarter of the input price; cache writes free or with a surcharge by model | Repeated prefixes in prompts, such as system instructions and documents | Yes |
| Bedrock provisioned throughput | Model unit per hour, without commitment or with a one- or six-month commitment at lower rates | Reserved throughput for one model, used or not; mandatory for custom models with full fine-tuning | Yes |
| Bedrock customisation | 1,000 tokens or hours for fine-tuning, month per stored custom model, model unit per minute for imported models | Every trained or imported model | Yes |
| Bedrock Guardrails, Knowledge Bases, Agents | Per 1,000 text units for policies; the vector database, embeddings and model invocations behind them at their rates | Enabled policies, queries, agent steps | Yes |
| SageMaker instances | Second per instance type and component: notebook, Studio, training, processing, batch transform, real-time and asynchronous endpoints | The instances that run, whether code executes or not; endpoints with several instances per instance | No |
| SageMaker Serverless Inference | GB-second of configured memory per invocation, plus provisioned concurrency per GB-second | Invocations and their duration, with no idle instance | Yes |
| SageMaker storage and data | GB-month for notebook volumes and model artifacts in S3; Feature Store per GB-month and request; data transfer | Volumes, artifacts, features | Yes |
| Free allowances | None on Bedrock; SageMaker AI 250 hours of ml.t3.medium for notebooks and further quotas in the first two months | New accounts | Yes |
| Purchase option | SageMaker Savings Plans over one or three years on instance usage; Bedrock provisioned throughput with commitment | Steady baseline | Yes (AWS Savings Plans) |
On Bedrock the model is the pricing dimension that sets everything: a large model costs several times a small one of the same family per token, and within a model an output token costs a multiple of an input token, four times on Amazon Nova Pro. Context is input: every invocation pays the whole prompt with system instruction, history and documents anew, and prompt caching lowers the price of repeated prefixes to a quarter. Batch halves both rates for jobs that tolerate a response time of hours; the Flex and Priority service tiers move the rate for the same tokens down or up. Provisioned throughput replaces token billing with model units per hour, with commitments over one or six months at lower rates, and is the only form in which a fully fine-tuned custom model runs. Cross-region inference spreads invocations across the regions of a geography or worldwide, with price tables of its own for the global variant; the prices per model and region are in the price list, and in Frankfurt the offer lists the models of Amazon, Meta and Mistral AI.
On SageMaker AI the instance-second per component counts: a notebook pays its instance as long as it runs, a training job the instances of its duration, an endpoint its instances around the clock until it is deleted. The same instance type costs the same in every component, and the surcharge over EC2 is the price of the management. Serverless Inference replaces the endpoint with GB-seconds of configured memory per invocation, with no idle, with cold starts; provisioned concurrency keeps capacity warm and costs per GB-second for it. The free allowance of the first two months covers small notebooks, not endpoints in production.
How it appears in your billing data
In the Cost and Usage Report the Bedrock lines carry the product code AmazonBedrock, and the usage type names model, token kind and tier: EUC1-NovaPro-input-tokens and EUC1-NovaPro-output-tokens for on-demand, EUC1-NovaPro-input-tokens-batch for batch, EUC1-NovaPro-cache-read-input-token-count for cache reads, EUC1-NovaPro-input-tokens-flex and EUC1-NovaPro-input-tokens-priority for the service tiers; the unit is 1,000 tokens, the prefix the region code, EUC1 for Frankfurt. Provisioned throughput appears in model-unit hours, imported models as custom model units per minute. lineItem/ResourceId names nothing on on-demand that distinguishes an application: all invocations of a model from one account fall into one line, and attribution to applications or customers needs application inference profiles with tags, whose ARN then appears as the resource, or separate accounts.
SageMaker lines carry the product code AmazonSageMaker, the operation RunInstance and the usage type made of component and instance type: EUC1-Notebk:ml.t3.medium for notebook instances, EUC1-Studio:JupyterLab-ml.t3.medium for Studio, EUC1-Train:ml.m5.xlarge for training, EUC1-Processing:ml.m5.xlarge, EUC1-Tsform:ml.m5.xlarge for batch transform, EUC1-Host:ml.m5.xlarge for real-time endpoints, EUC1-AsyncInf:ml.m5.xlarge for asynchronous ones, EUC1-ServerlessInf:Mem-2GB for Serverless Inference in seconds per memory size; lineItem/ResourceId names the endpoint, the training job or the notebook. The model artifacts sit under AmazonS3, the logs under AmazonCloudWatch.
What the reports do not deliver is the price per request or per user: Bedrock sums tokens per model, SageMaker seconds per endpoint, and only the application that counts tokens and invocations per customer turns them into unit costs; how that is done is described in AI cost allocation. The report's structure is described in the AWS Cost and Usage Report. In the FOCUS export, ServiceName carries Amazon Bedrock or Amazon SageMaker and x_ServiceCode AmazonBedrock or AmazonSageMaker; tokens and instance-seconds are ChargeCategory Usage, provisioned throughput with commitment and SageMaker Savings Plans PricingCategory Committed (FOCUS columns).
Levers
Configuration: on Bedrock the model is the first lever, because a smaller model of the same family is several times cheaper per token and suffices for many tasks; prompt length the second, because every invocation pays the whole context, and prompt caching the third for repeated prefixes. Batch for everything not needed at once, Flex for background load; the method is in caching, batching and routing and tokens, requests and context. On SageMaker, deleting is the first lever: notebooks with a schedule for idle time, endpoints deleted after the experiment, and Serverless Inference for models with rare invocations; the instance type the second, because GPU instances cost six times a CPU instance per hour and training and inference often run on smaller types.
Purchase option: for steady SageMaker load, SageMaker Savings Plans lower the instance-second over one or three years per component; for steady Bedrock load, provisioned throughput with commitment lowers the model unit, but pays off only when the units are utilised, because they cost per hour whether tokens flow or not (commitment strategy).
Architecture: Bedrock for applications that call a foundation model, SageMaker for custom models and training; an endpoint around the clock only for models with steady traffic, otherwise serverless or batch transform. What the models cost across providers is in model pricing and rate cards, and the dimensions across all three clouds in AI and machine learning services; the comparison of billing with Azure OpenAI is drawn in the article Azure OpenAI and Bedrock billing.
Where it goes wrong
- The context grows with every turn. A chat that sends the whole history pays it as input on every invocation. The input tokens per invocation in the metrics show it; summarising the history and prompt caching lower it.
- The large model for the simple task. Classification and extraction run on Nova Micro at a fraction of Nova Pro. Routing by task is the lever.
- Provisioned throughput without load. A model unit costs per hour whether tokens flow or not. The token line against the unit's throughput shows the utilisation; below a third, on-demand is cheaper.
- The endpoint from the experiment is still running. A real-time endpoint pays its instances until it is deleted.
Host:lines without invocations inInvocationsare the clue. - The notebook runs over the weekend. A GPU notebook instance pays per second whether anyone types or not. Lifecycle configurations stop notebooks after idle time.
- Batch transform would have been cheaper. An endpoint that scores a million records once a night pays 24 hours for one hour of work. Batch transform pays only the job.
- The invocations cannot be attributed to a team. Bedrock tokens fall into one line per model. Application inference profiles with tags or an account per application separate them.
Example
List prices in USD for Frankfurt (eu-central-1), read on 2026-10-21 from the AWS price list: Amazon Nova Pro costs on-demand USD 0.00105 per 1,000 input tokens and USD 0.0042 per 1,000 output tokens, in batch USD 0.000525 and USD 0.0021, for cache reads USD 0.0002625 per 1,000 tokens, in the Flex tier USD 0.000525 and USD 0.0021 and in the Priority tier USD 0.0018375 and USD 0.00735; Amazon Nova Micro USD 0.000046 and USD 0.000184 per 1,000 tokens. On SageMaker AI an ml.m5.xlarge (4 vCPU, 16 GiB) costs USD 0.276 per hour in every component, an ml.g5.xlarge with one GPU USD 1.761, an ml.t3.medium USD 0.058; Serverless Inference USD 0.00004 per second with 2 GB and USD 0.00012 with 6 GB.
An application that sends 10 million input tokens and 2 million output tokens a month to Nova Pro pays on-demand 10.50 plus 8.40 = USD 18.90; the same tokens in batch USD 9.45, and with prompt caching for 60 percent of the input 6.30 plus 1.05 plus 8.40 = USD 15.75. The same load on Nova Micro costs 0.46 plus 0.37 = USD 0.83. A real-time endpoint on an ml.m5.xlarge costs USD 201.48 a month around the clock, on an ml.g5.xlarge USD 1,285.53; the same model with Serverless Inference at 2 GB and a million invocations of 200 ms each costs 200,000 seconds × 0.00004 = USD 8.00. A notebook ml.t3.medium that is never stopped costs USD 42.34 a month. The figures show that output tokens on Nova Pro cost four times the input and carry almost half the bill, that the small model cuts the bill twentyfold, and that a GPU endpoint around the clock costs more than hundreds of millions of tokens.
Sources
Read on 2026-10-21. Prices: Amazon Bedrock pricing, Amazon SageMaker AI pricing, AWS Price List, offer AmazonBedrock for eu-central-1, offer AmazonSageMaker. Billing data: Cost and Usage Report columns: lineItem, FOCUS 1.0 with AWS columns.