A sudden spike from $1k to $400k in Amazon Bedrock costs over a few days is extremely unlikely to be a billing bug. This represents a massive increase in actual token consumption that would need investigation. Here's what's likely happening and what you should do:
Why This Happens: With Amazon Bedrock's token-based pricing model, costs can spike dramatically and quickly. Unlike traditional compute where costs scale predictably with instance hours, a single misconfigured application can consume millions of tokens in minutes. Common causes include retry loops without proper backoff logic, prompts passing entire documents without truncation, batch jobs without concurrency controls, or an application error causing excessive API calls.
Why Detection Was Delayed: AWS Cost Explorer updates hours after consumption occurs, so by the time cost anomalies appear in billing dashboards, significant charges may have already accumulated. However, Amazon Bedrock does publish InputTokenCount and OutputTokenCount metrics to CloudWatch within minutes of model invocations, which provides the earliest cost signal available.
Immediate Actions:
- Go to the AWS Support Center (https://console.aws.amazon.com/support/) immediately to add correspondence to your existing case or open a new one if you haven't already
- Check your CloudWatch metrics for InputTokenCount and OutputTokenCount for the specific model during the spike period
- Review CloudTrail logs to identify which IAM principal (user or role) made the calls
- Temporarily restrict access to the model if you suspect unauthorized usage
- Examine your application logs for errors, retry loops, or unusual patterns
Investigation Steps: Look at your application code for issues like infinite retry loops, missing rate limiting, or logic errors that could cause repeated calls. Even if only one model shows usage, this could indicate a specific code path or feature using that model had an issue rather than credential compromise.
For Future Prevention: Implement CloudWatch alarms on token metrics with appropriate thresholds, enable model invocation logging for detailed request analysis, use Application Inference Profiles with cost allocation tags for better tracking, and consider AWS Budgets for cost-based monitoring with automated alerts.
Given the magnitude of this spike, work with AWS Support immediately through the Support Center to investigate the root cause and discuss your options.
Sources
Analyze Amazon Bedrock cost spikes with AWS DevOps Agent | AWS re:Post
How to Track and Limit Amazon Bedrock Usage by User | AWS re:Post
Track Amazon Bedrock Costs by Caller Identity with IAM Principal-Based Cost Allocation | AWS Cloud Financial Management