Bedrock custom models no Provisioned Throughput or active deployment references

Custom models bill for storage even when nothing serves them

When you fine-tune or continue pre-training a model in Bedrock, the result is stored as a custom model in your account. The worked examples on the Amazon Bedrock pricing page include “custom model storage per month ($1.95)” as a line item in its fine-tuning scenarios. To use the model, you then set up inference either by buying Provisioned Throughput or by deploying it for on-demand inference.

Experiments leave many models behind: each training run with a different dataset or hyperparameters becomes another stored model, and only one of them ends up being served.

Matching models to what serves them

List the custom models, then the two things that can serve one, and compare the ARNs:

aws bedrock list-custom-models \
  --query 'modelSummaries[].[modelName,modelArn]' --output table

aws bedrock list-provisioned-model-throughputs \
  --query 'provisionedModelSummaries[].[provisionedModelName,modelArn,foundationModelArn]'

aws bedrock list-custom-model-deployments \
  --query 'modelDeploymentSummaries[].[customModelDeploymentName,modelArn,status]'

A custom model ARN that appears in neither list is serving nothing.

How ZopNight decides a model is orphaned

ZopNight builds the same cross-reference. A model counts as in use if any Provisioned Throughput points at it, or if a custom model deployment for it is Active. A deployment that is still being created or has failed is not serving inference, so it does not protect the model. When neither applies, the model is flagged, provided it has a monthly storage price.

Why a model might not be flagged

If either list call failed during discovery, the cross-reference is incomplete. A positive match found anyway still counts, but a model is never marked unused from partial data; it waits for a full pass. A model with no storage price in ZopNight’s data is skipped rather than shown at $0.

One storage fee per model

saving = custom model storage fee per month
cost after fix = 0

The saving per model is small; it adds up in accounts that fine-tune often.

Cleaning up unused custom models

1. Confirm with the ML team that no upcoming launch or evaluation needs the model.
2. Keep the training data and job configuration so the model can be rebuilt if needed.
3. Delete it with aws bedrock delete-custom-model --model-identifier and the model ARN.
4. Add a cleanup step to fine-tuning pipelines so losing candidates are deleted after evaluation.