Amazon SageMaker
SageMaker
You need managed tools for preparing data, training, evaluating and operating machine learning models.
Reach for it when
- Training and hosting a custom model with a reproducible evaluation workflow.
- Managing an ML lifecycle that needs more control than simply invoking a hosted foundation model.
Do not reach for it when
- Choosing a chatbot API when custom training is unnecessary.
- Skipping dataset quality, evaluation or monitoring because training infrastructure is managed.
Alternatives, and how to choose
| Service | Pick it instead when |
|---|---|
| Bedrock | Choose managed foundation-model access when that matches the task. |
| Your own compute workflow | Evaluate the cost and control tradeoffs of managing more infrastructure yourself. |
How you pay
- The model
- Selected notebooks, training jobs, hosting capacity, storage and other features affect cost.
- The line item that surprises people
- Idle notebooks and persistent endpoints can accrue charges outside active experiments.
What trips people up
- Good training metrics do not guarantee behavior on deployment data.
- Separate training, evaluation and production data access; inspect the offering and feature scope you actually use.
Verify the live service
This page is a concept reference. Cost models are qualitative; confirm the current offering, Region and pricing before deploying.