Three horizons cover a runaway Bedrock bill, and the guardrail layering uses all of them: a CloudWatch alarm on tokens in minutes, an anomaly monitor up to a day later, a budget across the month. This card is the middle one.
One scope detail matters on Bedrock. Cost Anomaly Detection leaves AWS Marketplace charges out, and third-party foundation models on Bedrock bill under the AWS Marketplace billing entity. Those models are the documented exception and are monitored. Other Marketplace charges need a cost budget instead.
Flash card
AWS Cost Anomaly Detection: machine-learning detection of unusual spend against a baseline learned from the account's own history, offered as a feature within AWS Cost Explorer at no charge.
- A monitor covers one of four dimensions: AWS services, linked account, cost allocation tag, or cost category. The AWS services dimension evaluates every service in the account, so there is no Bedrock-only monitor.
- Per-feature reporting comes from a cost allocation tag monitor, once Bedrock application inference profiles carry the tags.
- An AWS managed monitor tracks every value in its dimension independently, up to 5,000 of them. A customer managed monitor aggregates up to 10 values you select, or a single value for a cost category.
- Alert thresholds are absolute dollar impact or impact percentage, and two can be combined with AND or OR. Individual alerts need an Amazon SNS topic; the daily and weekly summaries go by email.
- Detection reads Cost Explorer data about three times a day, so an anomaly can take up to 24 hours to surface. A newly monitored service needs 10 days of history first.
Pick it when
Pick it over a budget when the signal you want is that spend looks unusual for this account, rather than that spend passed a number someone set. Pick it over a CloudWatch alarm when the departure should be measured in dollars rather than tokens or invocations.
It's the wrong answer when
It is the wrong answer for a same-minute reaction to a retry storm; that calls for a CloudWatch alarm on Invocations or token counts. It is wrong again for enforcing a hard ceiling, which needs a budget action, a service quota, or an application rate limiter, and for a feature with fewer than 10 days of spend history behind it.