Generative AI (GenAI) on AWS can be paid for in several ways, and each one trades cost against speed, availability or commitment. The exam gives a workload, such as a nightly batch job, a steady high-volume app or an unpredictable prototype, and asks for the most cost-effective option that still meets the requirement. This lesson explains how Amazon Bedrock pricing works and which option suits which workload. It gives no prices, because they change: learn the patterns instead.
On-Demand, Paying per Token
By default, Bedrock charges per token: you pay for the input tokens you send and the output tokens the model generates, with no commitment. That makes on-demand the natural fit for new, low or unpredictable traffic, because nothing is charged when nothing is used. Longer prompts and longer answers cost more and take longer, so trimming them saves money on every call.
Batch Inference for Work That Can Wait
Batch inference processes many prompts in one asynchronous job. You put the inputs in Amazon S3, submit the job, and collect the outputs from S3 when it finishes. For supported models, batch is priced lower than on-demand. The trade-off is time: results arrive when the job completes, not in real time. That suits overnight summarisation, bulk classification and dataset preparation, but never a live chat.
Provisioned Throughput for Steady, Heavy Use
Provisioned Throughput reserves a fixed level of capacity for a model, measured in model units, and is billed hourly whether you use it or not. You can choose no commitment, 1 month or 6 months; longer terms lower the hourly price. It suits predictable, high-volume workloads that need consistent performance. It's poor value for spiky or low traffic, because idle hours are still billed.
Service Tiers
For on-demand requests, Bedrock offers service tiers chosen per request:
| Tier | What it's for |
|---|---|
| Priority | Fastest responses for customer-facing workloads, at a premium price |
| Standard | The default for everyday tasks |
| Flex | A discount for work that can tolerate longer processing |
| Reserved | Reserved capacity for mission-critical apps, billed monthly for 1- or 3-month terms |
Other Levers on Cost and Availability
- Prompt caching: when many requests repeat the same long context, such as a document or system prompt, caching lets the model reuse it. Cached tokens cost less and responses come back faster.
- Cross-Region inference routes requests across several AWS Regions to handle bursts and improve availability, at no extra routing charge. A geographic profile keeps processing within an area such as the US or EU, which helps with data residency rules. A global profile can route to any supported Region and costs less, but it gives up that geographic boundary.
- Regional coverage: not every model is offered in every Region, which can limit choices or push you towards cross-Region inference.
- Custom models add costs: training is charged (for example, per token processed during fine-tuning) and each custom model has a monthly storage fee. Running it then needs Provisioned Throughput or a custom model deployment for on-demand use.
- Model choice: a smaller model, or a distilled one, costs less per token. Use the smallest model that meets the quality bar.
On the Exam
| Workload | Usual answer |
|---|---|
| Unpredictable or low traffic | On-demand |
| Large job, results needed later | Batch inference |
| Steady, high, predictable volume | Provisioned Throughput |
| Same long context in many requests | Prompt caching |
| Bursts plus data must stay in the EU | Geographic cross-Region inference |
| Fastest replies without reserving capacity | Priority tier |