Amazon Bedrock Pricing Explained: On-Demand, Batch and Provisioned Throughput

1–2 minutes

Updated

Short answer: most teams should start with on-demand pricing, move offline workloads to batch inference, and consider provisioned throughput only for steady, high-volume traffic with a predictable model.

The pricing options

OptionHow you payGood for
On-demandPer input and output token, no commitmentPrototypes, variable traffic, most production apps at first
Batch inferencePer token, at a discount versus on-demand for supported modelsOffline jobs: classification, summarization, evaluations
Provisioned throughputPer model unit per hour, with or without a term commitmentSteady high volume, custom models
Prompt cachingCached input tokens billed at a lower rate on supported modelsLong, repeated system prompts and documents

What actually drives the bill

  • Output tokens usually cost several times more than input tokens. Long answers are expensive answers.
  • Context size. Sending a whole document on every call multiplies input tokens; RAG and caching reduce it.
  • Model choice. Smaller models can be an order of magnitude cheaper for routing, extraction and classification.
  • Retries and agents. Agent loops call the model many times per user request.

Estimate before you build

estimate.py

monthly_cost = requests_per_month * (
    avg_input_tokens  / 1_000_000 * input_price_per_million +
    avg_output_tokens / 1_000_000 * output_price_per_million
)

Take the prices from the Bedrock pricing page for your model and region. They change often, which is why this site keeps a dated tracker instead of hard-coding them in articles.

Set maxTokens on every call. Without it, a model can return far more output than you need, and output is the expensive side.

Track it per application

Application inference profiles let you tag model usage per application or team, so Cost Explorer can split the Bedrock bill instead of showing one line.

Coming in this hub: a token cost calculator and a monthly benchmark of cost per task across Bedrock models.

Run the numbers for your account

Use the Bedrock token cost calculator with your own traffic and prices.

Open the calculator →

Joniel Gallo · Cloud & AI Engineer

I build these labs in a real AWS account and publish the code. Need a second opinion on your architecture or your AWS bill? Work with me.

One tested AWS & AI lab a week

Real costs, working code and what changed in Bedrock this month. Subscribe and get the free AWS Permission Debugger (PDF).