Skip to main content

Pricing

Key Takeaways​

Amazon Bedrock provides three core pricing and purchasing options for foundation model inference: On-Demand (pay-as-you-go per token/image), Batch Mode (asynchronous processing via Amazon S3 with a 50% cost discount), and Provisioned Throughput (reserving dedicated Model Units for consistent throughput).

+---------------------------------------------------------------------------------------+
|                      AMAZON BEDROCK PRICING & INFERENCE MODES                         |
+---------------------------------------------------------------------------------------+
|                                                                                       |
|   1. ON-DEMAND (Pay-As-You-Go)                                                        |
|      * Billed per 1,000 / 1,000,000 Input & Output tokens (or per generated image)    |
|      * Zero upfront commitment; ideal for bursty, unpredictable, or dev workloads     |
|                                                                                       |
|   2. BATCH MODE (50% Cost Savings)                                                    |
|      * Asynchronous bulk processing: S3 Input File ---> Bedrock ---> S3 Output File   |
|      * 50% discount vs. on-demand; ideal for non-real-time offline jobs               |
|                                                                                       |
|   3. PROVISIONED THROUGHPUT (Reserved Capacity)                                       |
|      * Purchases dedicated Model Units (MUs) on 1-month or 6-month commitments        |
|      * Guarantees max tokens/min throughput and zero throttling                       |
|      * MANDATORY requirement to run Custom / Fine-Tuned / Imported Models             |
+---------------------------------------------------------------------------------------+

When optimizing costs, note that inference parameters like Temperature, Top-K, and Top-P do not affect pricing—the primary cost driver is total input and output token volume.

Main Discussion​

The Three Bedrock Inference Pricing Modes​

+----------------------------------------------------------------------------------------------------+
|                               INFERENCE PURCHASING OPTIONS MATRIX                                  |
+--------------------------+-----------------------+--------------------+----------------------------+
| Inference Mode           | Billing Unit          | Commitment Terms   | Primary Architectural Fit  |
+--------------------------+-----------------------+--------------------+----------------------------+
| On-Demand                | Per Input/Output      | None               | Interactive web chatbots,  |
|                          | Token & Image Gen     | (Pay-as-you-go)    | prototyping, variable load |
+--------------------------+-----------------------+--------------------+----------------------------+
| Batch Inference          | 50% of On-Demand      | None (Job-based)   | Offline document analysis, |
|                          | Token Rates           |                    | nightly report summaries   |
+--------------------------+-----------------------+--------------------+----------------------------+
| Provisioned Throughput   | Hourly rate per Model | 1-month or 6-month | Production SLAs, high QPS, |
|                          | Unit (MU) committed   | (or no-commit 1 MU)| custom/fine-tuned models   |
+--------------------------+-----------------------+--------------------+----------------------------+
+-----------------------------------------------------------------------------------+
| BATCH INFERENCE ASYNCHRONOUS ARCHITECTURE (50% DISCOUNT)                          |
|                                                                                   |
|  [ Amazon S3 Input Bucket ] ---> Contains `input_prompts.jsonl`                   |
|              |                                                                    |
|              v (Trigger Async Batch Job)                                          |
|  [ Amazon Bedrock Batch Engine ]                                                  |
|  * Processes requests asynchronously with 50% discount                            |
|              |                                                                    |
|              v (Write Aggregated Completions)                                     |
|  [ Amazon S3 Output Bucket ] ---> Generates `output_completions.jsonl`            |
+-----------------------------------------------------------------------------------+
  • On-Demand Pricing: Billed independently for input tokens and output tokens. Output tokens generally carry a higher price per token due to the computational cost of sequential autoregressive token generation.
  • Batch Mode: You supply a JSONL file in Amazon S3 containing bulk prompts. Bedrock processes them asynchronously, outputs a single result file back to S3, and applies a flat 50% discount compared to standard on-demand token rates.
  • Provisioned Throughput: Designed for guaranteed capacity and strict latency SLAs. Crucial Rule: To run inference on custom fine-tuned models, you cannot use on-demand billing; you must purchase Provisioned Throughput.

The Cost Spectrum of Model Customization & Adaptation​

Adapting foundation models to business use cases follows a progressive cost and compute curve:

Engineering Cost: Prompt Engineering≪RAG<Instruction Fine-Tuning≪Domain Adaptation (Continued Pre-training)\text{Engineering Cost: } \text{Prompt Engineering} \ll \text{RAG} < \text{Instruction Fine-Tuning} \ll \text{Domain Adaptation (Continued Pre-training)}

+-----------------------------------------------------------------------------------+
| MODEL ADAPTATION COST & COMPLEXITY HIERARCHY                                      |
|                                                                                   |
|  [ Domain Adaptation / Continued Pre-training ]  $$$$$ (High Compute & Unlabeled) |
|  ----------------------------------------------                                   |
|  [ Instruction-Based Supervised Fine-Tuning ]    $$$   (Moderate GPU & Labeled)   |
|  ----------------------------------------------                                   |
|  [ Retrieval-Augmented Generation (RAG) ]        $$    (Vector DB & Storage Costs)|
|  ----------------------------------------------                                   |
|  [ Prompt Engineering & System Instructions ]    $     (Zero Training Compute)    |
+-----------------------------------------------------------------------------------+
  • Prompt Engineering ($): Zero additional model training, storage, or infrastructure overhead. You pay only standard inference token charges for prompt optimization.
  • Retrieval-Augmented Generation / RAG ($$): Base model weights remain untouched. Adds operational costs for document storage in Amazon S3, embeddings generation, and vector database capacity (such as Amazon OpenSearch Serverless OCUs).
  • Instruction-Based Fine-Tuning ($$$): Modifies model weights using labeled input-output pairs. Incurs one-time GPU training costs plus mandatory Provisioned Throughput for inference hosting.
  • Domain Adaptation / Continued Pre-training ($$$$$): Massive pre-training over large volumes of unlabeled raw domain text to expand foundational vocabulary and domain reasoning. Highest computational and financial expense.

Cost Optimization Levers & Token Management​

+-------------------------------------------------------------------------------+
| BEDROCK COST OPTIMIZATION LEVERS                                              |
|                                                                               |
|  1. INPUT TOKEN PRUNING:                                                      |
|     * Strip redundant boilerplate and irrelevant conversational history       |
|                                                                               |
|  2. OUTPUT LENGTH CONSTRAINTS (`max_tokens`):                                 |
|     * Cap maximum completion tokens and enforce concise generation formats    |
|                                                                               |
|  3. RIGHT-SIZING FOUNDATION MODELS:                                           |
|     * Route short NLP/classification tasks to lightweight models (e.g., Nova  |
|       Micro/Lite or Titan Text) instead of expensive flagship reasoning models|
|                                                                               |
|  4. LEVERAGING BATCH MODE:                                                    |
|     * Shift non-urgent, offline bulk jobs to batch inference for 50% savings  |
+-------------------------------------------------------------------------------+

  • Inference Parameters vs. Cost: Adjusting Temperature, Top-P, or Top-K alters model creativity and token selection distributions, but has zero direct impact on pricing.
  • Token Volume is the Pricing Driver: The most effective lever for reducing Bedrock expenses is minimizing unnecessary input prompt tokens and capping generated output tokens.

Exam Guide​

Exam Tips​

  • Batch Mode Discount: Remember the exact savings figure: Batch inference provides a 50% discount compared to on-demand pricing for non-real-time workloads processed via Amazon S3.
  • Provisioned Throughput Mandate: If an exam question mentions deploying a customized, fine-tuned, or imported model for inference, you must select Provisioned Throughput (on-demand mode is not supported for custom models).
  • Provisioned Throughput is NOT a Default Cost Saver: Provisioned Throughput is designed for guaranteed capacity, rate-limit avoidance, and predictable SLAs, not general cost reduction. If utilized below capacity, it can be more expensive than on-demand.
  • Hyperparameters vs. Billing: Parameters like Temperature, Top-K, and Top-P do not change token pricing. Only total tokens consumed and model tier dictate cost.

Practice Test​

Question 1​

A media company needs to generate text summaries for 100,000 archived news articles every weekend. The summarization process is not time-sensitive, and results can be delivered within 24 hours. The engineering team wants to minimize inference costs on Amazon Bedrock. Which approach provides the highest cost savings?

  • A. Provision a dedicated Provisioned Throughput Model Unit with a 6-month commitment
  • B. Submit the documents as an asynchronous Batch Inference job reading and writing to Amazon S3
  • C. Increase the Temperature and Top-P inference parameters on an On-Demand endpoint
  • D. Perform supervised fine-tuning on an Amazon Titan model using labeled summaries
Correct Answer
  • B. Submit the documents as an asynchronous Batch Inference job reading and writing to Amazon S3
    • Explanation: Batch Inference on Amazon Bedrock processes non-real-time bulk workloads asynchronously via Amazon S3 at a 50% discount compared to standard on-demand pricing, making it the most cost-effective solution for non-urgent tasks.

Question 2​

A machine learning engineer has completed a supervised fine-tuning job on Amazon Bedrock to adapt an Amazon Nova model for medical billing classification. The engineer now needs to deploy the custom fine-tuned model for production inference. Which purchasing option is required?

  • A. On-Demand pricing per 1,000 tokens
  • B. Provisioned Throughput
  • C. Spot Model Units with automatic scale-in
  • D. AWS Free Tier on-demand allocation
Correct Answer
  • B. Provisioned Throughput
    • Explanation: Amazon Bedrock requires Provisioned Throughput to host and serve inference requests for custom, fine-tuned, or imported models. On-demand pay-as-you-go inference is supported only for base foundation models.