How to raise your Amazon Bedrock quotas

How do Bedrock quotas work?

Bedrock has no tiers. Each model has its own quotas in each Region, counted in tokens per minute, and some models add a requests per minute quota. Embedding models are the exception, since they’re limited by requests per minute.

The two endpoints count tokens differently, which matters when you plan capacity. bedrock-runtime counts input and output tokens against one quota, and output tokens burn it down at a model-specific rate. bedrock-mantle has separate input and output quotas.

Where do you see your real quota?

In the Service Quotas console, under Amazon Bedrock. AWS is clear that the published defaults aren’t a promise. Your account’s values can be lower depending on Region, account age, usage history and payment history.

How do you request an increase on bedrock-runtime?

1. Check the model first. If it’s in Legacy or Deprecated status, AWS won’t raise the quota, so move to the successor.
2. Look at the Adjustable column for the quota. If it says Yes, you can request a change through Service Quotas.
3. For token quotas, request the Cross-Region InvokeModel tokens per minute quota for your model.
4. The support team then offers to raise the on-demand tokens per minute and the cross-model daily tokens quotas along with it.

How do you request an increase on bedrock-mantle?

Not through the Service Quotas console, even though the quotas show up there. Open the AWS Support limit increase form, select Amazon Bedrock, and give:

• The endpoint, bedrock-mantle.
• The Region.
• The model.
• The quota name, input TPM or output TPM, and the value you want.

You can ask for input and output on the same model in one case. Approval depends on whether your existing usage justifies it, so include recent numbers from CloudWatch or Service Quotas.

Why do requests get turned down?

AWS says that due to demand, priority goes to customers whose traffic consumes their existing quota, and that a request may be denied otherwise. A new account asking for a large number before it has any traffic is the weak case. Send real traffic first, then ask with the graph.

A higher quota doesn’t grant model access, either. If you can’t invoke the model at all, that’s a separate problem, and AWS says to contact sales about it.

What are the alternatives?

Provisioned Throughput reserves capacity for a model. Batch inference suits work that can wait. Both have quotas of their own.

See how Bedrock’s rules compare with other vendors on our AI API rate limits page.