Cost

The Cost page shows what one Environment spent on model calls, on judges and on the simulated user of simulations. It needs no plan feature.

The Cost page

Choose a range of 24 hours, 7 days (the default), 30 days or month to date. Five figures head the page: Total, LLM calls (what the model calls cost), Judge spend, Simulated user and Unpriced models, the number of models with no price. Below them, Group by switches the table between models, services and days. Each row shows the calls, the input, cached and output tokens, and the cost in USD. Unpriced models have no USD figure and are listed under the table.

Click a model to open Traces filtered to that model. A trace shows its own cost, and its Optimize button finds what to fix, cost included, and writes a fix prompt. Optimize needs the trace-optimization plan feature.

How a number is computed

Spans record token counts, and the page prices them each time it loads. The trace service returns the tokens for each model in the window. Within a trace, spans that report identical token counts count as one call, as when a gateway and the service behind it both record it. Prices come from the model catalog's price sheet in dollars per million tokens, with any override set by a platform admin taking precedence. Uncached input, cache-read, cache-write and output tokens each have their own price. Cache-read and cache-write tokens are part of the input tokens and are priced once.

Judge spend is the sum of the judge cost recorded on each trace evaluation in the window, manual and online. It is counted apart from model spend, so nothing is added twice. A catalog price change applies to model spend at once. Judge costs and the costs stored on finished runs were fixed when they ran and do not change.

Simulated user is what the simulated user's model calls cost: the user that plays SDK simulations live and the runs against endpoint and AWS AgentCore agents, on your organization's Model. Each call is priced when it is made and kept per Environment and UTC day, so a range counts every whole UTC day it touches; a 24-hour range that crosses midnight UTC counts both days. It is part of Total. A sandbox run's simulated user is not in it.

The budget card

The Online-eval budget card shows the online rule's judge spend today against its daily budget, whether the rule is enabled, disabled or stopped by budget, and how many traces sampling skipped. A rule stopped by its budget resumes at UTC midnight and does not score the traces it skipped. Change budget sets a new daily amount in USD, or Unlimited. If the new amount is above today's spend, a paused rule resumes on its next check, within about a minute. View online rule opens the rule. To be warned before the budget runs out, create an alert rule on Judge spend (see Alerts). Details are in Online rules.

Estimate a workload

The cost estimator needs the cost-calculator plan feature; the Cost page has no link to it. From a description of your agent it builds a use case, a workflow graph with tokens per node and scenarios with probabilities, then estimates the cost per conversation and per month with a breakdown and a confidence score. Reports are saved per Environment. A simulation run can verify a report, and its observed cost per conversation is recorded beside the prediction. See Launch and read a run.

Limitations

• The trace list sorts by status and latency only, so it cannot sort by cost. The trace API can sort by tokens, the closest proxy.
• There is no per-user cost. The user rollup carries token counts and no dollars, so the Users page shows tokens only.