Know what it costs

Control Room → Budgets. The organisation has a monthly budget, with what has been spent against it so far this period. Spend resets each period.

The budget warns at 80% and blocks calls once spend reaches 100%. The warning goes to the people named on the alert, here by email.

New Alert adds another budget beside the global one, with its own amount, period and warning threshold. You do not need to add one now.

Fine-tuning has its own ceiling: a job whose estimated cost would breach it is refused when it is submitted. No ceiling is set on this demo; the provider sets the value.

Click Cost breakdown to open Cost Analytics.

Total LLM spend for the window, with the number of requests it bought. The window is the last seven days; the picker above changes it.

Average cost per LLM call, next to the input and output tokens processed. This is the number to watch when a team switches to a larger model.

Each bar is a six-hour window of spend. A step up in the bars is where to look first: a new agent, a longer prompt, a model change.

Click LLM Usage.

The same calls, counted: total requests, the share that failed, average latency. Cost and volume are two tabs over one set of metered calls.

Control Room → Monthly Usage. GPU-hours used this month against the cap set by your operator, measured from the deployment cost ledger. In this release the cap is advisory: a deployment over it is not stopped.

The estimated infrastructure cost for the month. CPU and memory hours are estimates from each deployment's size; GPU-hours are measured.