Already on Bedrock.
Bedrock gives you the tokens and the bill. What it does not give you is who spent them, what they produced, and which change is worth making. That is the part Valistry adds, read-only, in an afternoon.
What it reads
- CloudWatch, AWS/Bedrock
- Input, output, cache-read and cache-write tokens, and invocations, per model per day.
- Cost Explorer
- The daily charge for the Bedrock service, read beside the ledger and never used to fill it.
- Access
- A read-only IAM user: cloudwatch:ListMetrics, cloudwatch:GetMetricData, ce:GetCostAndUsage. Nothing else is called.
- Never
- No prompt, no completion, no tool argument. Those are not fields this source reads, so they cannot reach us.
Four changes that cut a Bedrock bill
Each is detected from your own rows and carries its evidence and the change to make.
Each is measured against the same baseline seven days on. Nothing counts as saved until it is.
- Prompt caching
- Bedrock publishes cache-read and cache-write tokens apart from input, and a cache read is a fraction of an input token. Where the same context is sent repeatedly and never cached, the finding prices the difference on your own rows.
- The model for the job
- A cohort of low-complexity calls on a premium model, priced on every model the card carries, with the exact tokens those calls used. The quality delta is measured where an outcome source is connected, and said plainly where it is not.
- Retries
- Third and later attempts on the same run, counted and priced. A cap is a setting, not a rebuild.
- Prompt size
- Calls sharing a prompt's shape that run at ten times the tokens of the leanest. The evidence is the same prompt observed running leaner — read from a hash of its shape, never its text.
What it cannot do, said up front
- Per key or per application
- CloudWatch's Bedrock metrics carry a model dimension and no key or application dimension, so a shared account cannot be split by key from this source. A gateway or your own telemetry can, and attribution rules do the rest.
- Provisioned Throughput
- Capacity bought by the hour and not served is the one waste nobody in this market shows. We read the commitment API and do not yet model it: naming that is better than a number we cannot stand behind.
- Batch inference
- Batch is discounted against on-demand and is not separable in the metrics we read. Where it is used, the bill check carries the difference.
- A model newer than the card
- A model the rate card has no entry for is priced from its family, and every screen says what share of spend that is. The card is refreshed by a person reading a diff, never automatically.
Beside the rest of the estate
Most estates are not only Bedrock. The same ledger takes Anthropic and OpenAI directly, the gateways in front of them, your coding seats, and an export for anything else.
So a Bedrock figure sits beside the rest at the same grain, under the same owners, against the same outcomes.
The connector page lists the exact addresses it calls and the steps to connect it.