# Cost & usage

The Cost dashboard — LLM spend over time, broken down by model, user, organization, and billing component, plus the most expensive sessions and per-tool token volume.

The **Cost** dashboard tracks where your LLM spend goes and catches runaway usage early. For the arithmetic behind each number, see [Cost calculation](/docs/platform/cost-calculation.md).

## How it's populated

Brizz uses the cost the provider reported when there is one. Otherwise it prices the call itself, which needs the model and provider recognized and both rates and token counts available — if any is missing, the call isn't priced. The tool and prompt-size charts measure something different: estimated **token volume**, not money.

## How to read it in the dashboard

- **Cost over time** — spend against session count. Spend rising while sessions stay flat points at a per-session regression. The two use different clocks (spend by call time, sessions by ingestion time), so treat a divergence as a lead, not a measurement.
- **Cost by model** — where a model swap shows up.
- **Cost by user** / **Cost by organization** — top spenders, for attributing cost to a customer or team.
- **Cost by component** — input, cache read, cache write, output, and tools. **Unattributed** fills the gap when those come to less than the total.
- **Cache efficiency** — the share of input tokens served from cache. Cached input bills less, so a low share is worth investigating; the chart shows the current share, not how much of the rest could be cached.
- **Most expensive sessions** — sessions whose *start* is in range, ranked by full cost. A session that began earlier won't appear even if its spend landed inside the window.
- **Tools by token contribution** and **Tool size over time** — estimated tokens per tool, not per-tool spend.
- **System prompt size over time** — average estimated system-prompt tokens per agent. Markers show when the prompt changed; they carry the change, not a release identity.

Session filters and the time range scope the charts — except change-event markers, which follow only the service and date range.

## How to act on it

1. **Look at cache efficiency early.** Moving input into the prompt cache is often cheaper and faster than shortening prompts.
2. **Split growth from regression.** Flat sessions with rising spend points at per-session usage — a longer prompt, a chattier tool, a bigger model. Remember the clock difference: delayed ingestion produces the same shape.
3. **Line up a jump with a change event.** A step in the prompt-size chart often coincides with a step in the bill — a correlation worth chasing.
4. **Check schema tokens on the busiest tools.** If your provider re-sends definitions each call, trimming one pays out on every turn. See [Tools](/docs/platform/tools.md).
5. **Open the expensive sessions.** Outliers are often a retry loop or a runaway agent rather than steady-state cost.

## Availability

On by default. If **Cost** isn't in your navigation, [contact us](mailto:support@brizz.ai). Cost by organization needs [identified organizations](/docs/instrument/identify-organizations.md).

## See also

- [Cost calculation](/docs/platform/cost-calculation.md) — billing classes, rate sources, and how a session's cost is derived.
- [Tools](/docs/platform/tools.md) — per-tool token volume, including estimated schema overhead.
- [Identify users](/docs/instrument/identify-users.md) — what makes the per-user breakdown work.