How to see what each feature cost in Claude Code tokens

Why “per feature” is the hard part

Claude Code can show what a run used in tokens. What it can’t tell you on its own is what a feature cost, because a feature is usually several runs across a few sessions, and the usage isn’t grouped that way.

So the job is to tie tokens to the unit you care about: the task or the feature. Once the tokens hang off a task, “what did this cost” becomes a number you can read instead of a guess.

The free ways first

Before any paid tool, three free ones tell you something about spend. Each is useful, and each stops short of per-feature cost.

The /usage command

Type /usage in a session and the Session block shows that session’s token usage and a dollar figure. On an API account the figure is an estimate, computed locally at list price; the Claude Console is the authoritative bill. On a Pro or Max subscription the session cost isn’t relevant for billing, so you see plan usage bars and activity stats instead. Either way it is per session, and it resets on /clear.

ccusage

ccusage is a free command-line tool that reads Claude Code’s local log files and reports usage by day, month and session, with 5-hour billing-window blocks. Run it with npx ccusage@latest, then ccusage daily or ccusage session. It answers “what did today cost” well; its unit is the day or the session.

OpenTelemetry export

For teams, set CLAUDE_CODE_ENABLE_TELEMETRY=1 and point Claude Code at a collector, and it emits claude_code.token.usage and claude_code.cost.usage (the cost is an approximation). You can slice those by user, model, skill or subagent on a dashboard.

All three answer “what did I spend” by session, day or model. None answers “what did this feature cost,” because none ties spend to the task you were building. That is the gap the rest of this guide fills.

Four numbers to capture per run

A run bills in four buckets, and you want all four to get the cost right:

• Input and output tokens, the fresh ones.
• Cache write and cache read tokens, from prompt caching.

Cache reads usually dominate a long session while barely moving the bill, which is why a feature often costs far less than its raw token count suggests. If you have a run’s four counts, drop them into the cost calculator to get the API-equivalent cost.

Reading it per task in Hatchlog

Hatchlog measures the Claude Code tokens each task spent and shows them on the task’s Tokens block, with the API-equivalent cost beside them. The Overview rolls the same numbers up: the most expensive tasks over a date range, and the cost for the whole project. That’s cost per feature without adding anything up by hand, and without a dashboard to stand up. For what’s measured and what the labels mean, see token cost per task.

Where to get each number

You want	Where to look
This session’s spend	The /usage command
Spend by day or week	ccusage
Cost of one task or feature	The task’s Tokens block
Most expensive tasks over time	The Overview, by date range
Cost from a run’s raw counts	The cost calculator

Token cost is measured for Claude Code only. Cursor and Codex can use the same board, but their work has no cost figure.