Claude Code usage analytics: what is worth measuring

The distribution is the first finding

Before any particular metric, understand the shape. Agent spend follows a power law: a small number of long, unsupervised sessions account for most of a month.

The one average worth holding is the external one. Anthropic reports that across enterprise deployments Claude Code averages around $13 per developer per active day and $150 to $250 per developer per month, with 90% of users staying under $30 per active day, as of August 2026. If your own number is an order of magnitude off that in either direction, the interesting question is why, and it is usually answered by the next four metrics.

The four metrics worth tracking

1. Cost by repository

The most actionable cut available, and the one no organisation-level API can give you, because repository is only recorded locally in the working directory of each request.

A repo that costs three times another for comparable work is telling you something concrete: it is hard to navigate. In practice that means one of four things, in roughly this order of frequency: missing or stale project memory, a structure the agent has to rediscover every session, a test suite slow enough that the agent burns turns waiting on it, or generated files large enough that reading one blows the context window.

The fix is a change to the repository, not to how people prompt, and every human who works there benefits from it too.

2. Cache hit ratio

Cache reads divided by total input tokens. Healthy sustained sessions sit high, usually 70 to 90 percent, because the conversation prefix is stable and re-read at a tenth of the base input rate.

A ratio that falls over time is context churn. The two common causes are a project memory file being edited constantly, and gaps longer than the cache lifetime: on a subscription the prompt cache lives about an hour, and it drops to five minutes once you are drawing on usage credits or working on an API key, so a lunch break costs a full context reprocess on the first message back.

3. Edit acceptance rate

How often a proposed edit is accepted rather than rejected. This used to be a guess and is now a field in two places.

# Analytics API, per user per day
rate = accepted / (accepted + rejected)
  over tool_actions.{edit_tool, multi_edit_tool, write_tool, notebook_edit_tool}

# OpenTelemetry, live
sum(rate(claude_code_code_edit_tool_decision_total{decision="accept"}[1d]))
  / sum(rate(claude_code_code_edit_tool_decision_total[1d]))

It is the only routinely available signal about whether output quality is going up or down, and it moves before anyone files a complaint. A step change after a model switch, a new convention in project memory, or a plugin rollout is worth taking seriously the same week.

4. Cost per merged pull request

Total agent spend divided by pull requests merged. It is crude and it is the only metric on this list that connects spend to something delivered. Watch the trend, not the level: the level depends entirely on how big your PRs are.

If you have connected the GitHub integration, Anthropic attributes merged PRs to Claude Code and labels them claude-code-assisted, which gives you a denominator you did not have to build. The attribution is deliberately conservative: it matches added lines against session output over a window from 21 days before merge to 2 days after, ignores lines under four characters and generated files, and drops attribution entirely when a human rewrote more than 20 percent of the line. Treat the resulting count as a floor.

Metrics that look useful and are not

Where the money actually goes inside a session

Cost by repo tells you which project is expensive. The newer attribution data tells you which part of the agent is expensive, which is a different and often more surprising answer.

• Subagents. The query_source attribute splits main from subagent and auxiliary. Delegating a verbose task is usually cheaper overall, but a subagent that re-reads the codebase is not.
• Agent teams. Anthropic reports roughly 7x the tokens of a standard session when teammates run in plan mode, because each teammate carries its own context window.
• MCP servers. Attributed individually, and one chatty server can dominate a session that felt cheap.
• Skills and plugins. Attributed by name, so an over-eager skill shows up rather than hiding in the total.
• Effort level. Thinking tokens bill as output. The effort attribute makes the trade visible instead of theoretical.

Inside the CLI, /usage shows the same breakdown for the last 24 hours or 7 days on a Pro, Max, Team, or Enterprise plan, and flags any behaviour that accounts for 10 percent or more of recent usage, such as long context or cache misses. Press d or w to switch window. Those figures are computed from local session history on that machine, so they exclude other devices and claude.ai.

Turning a finding into a change

Questions people ask

Keep reading