---
title: Cost
description: Token and spend analytics across every runtime and model — breakdowns, trends, cache efficiency, attribution, anomaly detection and the optimizer.
keywords: LLM cost tracking, agent token usage, AI spend analytics, Claude Code cost, cache hit rate
eyebrow: Dashboard
---

# Cost

Where the money is going, at every level of detail you might need it.

## The breakdowns

**By runtime.** The first cut, and usually the surprising one. Which of your
agents is actually expensive.

**By model.** Per-model daily token and cost rollup. This is where you find out
you have been running a frontier model on a task a cheap one handles.

**By session.** Every session with its totals, sortable. The tail of this list
— a handful of sessions that cost more than the rest combined — is where the
value is.

**By skill and by tool.** Where a runtime attributes work to a named skill or
tool, cost follows the attribution.

**By team.** Team mappings let you roll individual keys or agents up to a group,
which is what makes an internal chargeback conversation possible.

## Trends and windows

Daily, weekly and monthly totals with the trend. One thing worth knowing: a
"cost window" has exactly one definition in ClawMetry, applied everywhere. The
number on Home, the number on this page and the number in `clawmetry status`
are the same query against the same rollups — if they ever disagreed, that
would be a bug rather than a subtlety.

## Cache efficiency

Prompt caching is the single largest lever on agent cost, and it is nearly
invisible without instrumentation.

- **Cache hit rate** over time
- **Cache trends** — is your caching getting better or worse as prompts drift?
- **Cache risk** — sessions where a small change would invalidate a large cached
  prefix

```bash
curl -s localhost:8900/api/efficiency/cache-hit-rate | jq
curl -s localhost:8900/api/usage/cache-trends | jq
curl -s localhost:8900/api/usage/cache-risk | jq
```

## Anomalies

Spend that does not look like your normal pattern gets flagged. This is a
statistical comparison against your own history, not a fixed threshold — a
$40 session is unremarkable on one fleet and an emergency on another.

```bash
curl -s localhost:8900/api/usage | jq '.anomalies'
curl -s localhost:8900/api/token-velocity | jq
```

→ [Anomaly detection](/docs/cost/anomalies/)

## Optimization recommendations

The routing advisor looks at what your sessions actually do and suggests where
a cheaper model would have produced the same result, plus where caching is being
left on the table.

```bash
curl -s localhost:8900/api/usage/optimization-recommendations | jq
curl -s localhost:8900/api/efficiency/routing-advisor | jq
curl -s localhost:8900/api/cost-optimizer | jq
```

→ [Reducing spend](/docs/cost/optimization/)

## Where the numbers come from

Cost is derived **at ingest**, once, from the token split and the model, using
a multi-provider pricing table. Deriving at read time would mean a change to the
pricing table silently rewriting history.

Where a runtime records real dollars — Cline, opencode, Pi, Copilot, Exo,
Goose for paid providers — those are used instead of an estimate, and the
session is marked `exact`. Where neither is available the session says
`unavailable` and shows no dollar figure at all.

The per-runtime detail is in [What each runtime exposes](/docs/runtimes/coverage/);
the full model is in [How cost is computed](/docs/cost/model/).

:::warning Do not compare against a tool that estimates
Cost tools that reconstruct spend from transcripts without cache accounting can
overstate by several times, because a cache read is priced very differently from
a fresh input token. If ClawMetry's number disagrees with another tool's, check
whether the other one is accounting for cache reads and writes separately.
:::

## The API

```bash
curl -s 'localhost:8900/api/usage'                        | jq
curl -s 'localhost:8900/api/local/models?since=2026-08-01' | jq '.rows'
curl -s 'localhost:8900/api/local/runtimes'               | jq '.rows'
curl -s 'localhost:8900/api/local/aggregates'             | jq '.rows[-14:]'
```