Wring Bedrock
Per-model token usage and estimated cost for Amazon Bedrock, minutes after it happens.
If you build on Amazon Bedrock, your daily cost sync is too slow to be useful: a prompt change that triples your token spend shows up in AWS billing a day or more later, by which time it has been running all night.
Wring Bedrock closes that gap. It reads Bedrock's own CloudWatch metrics from your account, so usage appears a couple of minutes after the calls happen — input and output tokens, invocations, and estimated cost, broken down per model.
What you see
- Tokens and invocations per model, so you can tell which model is driving the volume — not just that "Bedrock" got more expensive
- Estimated cost, per model and in total
- Today, the last 24 hours, or the last 7 days, with a chart over time
- Spend spikes — when a model's usage runs at three times its own recent baseline, the dashboard says so, marks where it started, and projects what the spike costs you if it keeps running. Ask Werner why hands the spike straight to Ask Werner with the context attached.
Cost here is an estimate
The numbers on this page are computed from token counts and public Bedrock list prices. That makes them fast, and slightly optimistic in your favour or against it depending on your setup: they don't know about prompt-cache reads, batch inference, or provisioned throughput.
Treat this dashboard as your early warning and per-model breakdown. Your cost insights and your invoice remain the authoritative numbers — those come from AWS's own billing data.
Enabling it
Wring Bedrock is an opt-in level: it reads CloudWatch metrics, which pure cost access doesn't include, so it ships as its own CloudFormation stack per AWS account — enable it on the accounts where you run Bedrock. The role it creates can do exactly two things: list which metrics exist and read their values.
That is narrower than it may sound:
- No Bedrock API access. The role cannot call Bedrock, list your models, or see any prompt or completion. It reads counters, not content.
- No logs. You don't need to turn on Bedrock model-invocation logging, and log contents remain explicitly denied to Wring.
- Read-only, revocable. No metric can be written, and deleting the stack ends the access.
Good to know
- Metrics come from the regions where you actually use Bedrock; adding a new region is picked up on its own.
- An account without the stack simply doesn't appear on the dashboard.
- Because the data is near-real-time rather than billing-grade, very recent minutes can still move slightly as CloudWatch fills in.
Open Bedrock usage in your dashboard, or write to [email protected] if you'd like help reading a spike.