Amazon Bedrock provisioned throughput: model units, commitment terms, pricing
How Amazon Bedrock Provisioned Throughput works: what a model unit buys, the three commitment terms, how it is billed and what AWS publishes about price.
For: Teams sizing or procuring reserved Bedrock capacity, including for customised models
Last verified How we verify
Amazon Bedrock serves most models on demand, priced per token. Provisioned Throughput reserves capacity for a specific model instead, sold in model units (MUs) and billed by the hour. This guide sets out the mechanics from AWS’s own documentation and pricing page, checked on 22 October 2026. The vendor profile is at Amazon Bedrock Provisioned Throughput.
What a model unit buys
AWS defines an MU by throughput for one specified model: the number of input tokens it can process, and the number of output tokens it can generate, across all requests in one minute. What one MU delivers differs by model, and AWS directs buyers to their account manager for what an MU specifies, per-MU pricing and limit increases.
Commitment terms
You choose a term when you buy:
| Term | What it means |
|---|---|
| No commitment | Delete the Provisioned Throughput at any time |
| 1 month | Cannot be deleted until the month is over |
| 6 months | Cannot be deleted until the six months are over |
The longer the commitment, the lower the hourly price. Billing continues until you delete the Provisioned Throughput, so a no-commitment purchase still costs money for every hour it exists.
Customised models need it
If you customise a model on Bedrock, AWS says you must purchase Provisioned Throughput to use it. A custom model is priced the same as the base model it was customised from. That makes Provisioned Throughput a requirement, not an option, for fine-tuned models on Bedrock.
What AWS publishes about price
The Bedrock pricing page publishes on-demand per-token prices and worked Provisioned Throughput examples rather than a full per-MU price list. One example on 22 October 2026: two Amazon Titan Text Express model units on a 1-month commitment at $18.40 per hour each, totalling $27,379.20 for a 31-day month. For other models, budget from a quote: the documentation points to the account manager for per-MU pricing.
Where you can buy it
AWS lists Provisioned Throughput availability by model and region. On 22 October 2026 the table named the regions us-east-1, us-west-2, ca-central-1, sa-east-1, eu-central-1, eu-west-1, eu-west-2, eu-west-3, ap-south-1 and ap-southeast-2, with availability varying by model. Under section 1.4 of the AWS Customer Agreement you may specify the regions where your content is stored. That is a residency commitment; the parent company remains US-based, which is a separate jurisdiction question.
Provisioned or on-demand?
The trade is the same as on other platforms: reserved, hourly-billed capacity against per-token prices on shared serving. See provisioned throughput vs pay-per-token for how to work out which is cheaper for your traffic, and the comparisons with Azure OpenAI PTU, Vertex AI Provisioned Throughput and Fireworks AI.