How to Check Claude Code Token Usage

Claude Code is unusually transparent about usage: it writes a JSONL log of every message, including token counts, to a hidden folder on your machine. The challenge is that the built-in slash commands are designed to help you stay within limits, not to give you a historical audit. If you want to know what you burned last Tuesday, or whether your Claude Max subscription is actually cheaper than API billing would be, you need to read the logs directly — and that is exactly what whoburnedmore does.

How do I check my Claude Code token usage?

The fastest path is a single command that requires nothing installed and no account:

$ npx whoburnedmore↳ scanning ~/.claude/projects/…↳ found 847 JSONL entries across 23 sessions   CLAUDE CODE USAGE  ────────────────────────────────────────  2026-06-14  input  2,341,200  output  189,400  2026-06-13  input  1,876,500  output  142,900  2026-06-12  input  3,104,800  output  267,100   MODEL BREAKDOWN  claude-sonnet-4-5       14.2M tokens  claude-opus-4           4.1M tokens   30-day total:  18.3M tokens   est. $61.40

The ~/.claude directory (or ~/.config/claude on some Linux systems) is where Claude Code writes its project logs. whoburnedmore uses the open-source ccusage library under the hood to parse these JSONL files, then aggregates them into daily, monthly, and session-level summaries — adding up the inputTokens and outputTokens fields from each entry.

What the built-in slash commands show

Claude Code ships three commands that give you partial usage visibility inside an active session. Understanding what each one covers — and what it misses — helps you decide when to reach for whoburnedmore:

1. 1

   /usage — current rate-limit window

   Shows how many tokens you have consumed in your current 5-hour or weekly window relative to your plan cap. Useful for knowing if you are about to hit a limit. Does not show cost or historical data.
2. 2

   /context — context-window fill

   Shows how full the current conversation context window is (expressed as a percentage). This tells you when the model will start dropping earlier messages, but it is a property of the current session, not a usage total.
3. 3

   /cost — session-level cost

   Shows the estimated cost for the current session only. Close the terminal and the number is gone. No persistence, no day-over-day comparison.

What does a daily token breakdown look like?

Token consumption varies wildly by day. A day spent on a large refactor with heavy tool-use loops burns far more tokens than a day of quick questions. The daily view makes those spikes visible 📊:

Jun 12 stands out — that is the day a large refactor ran 47 sequential tool-call loops. Each loop sent the full context back to the model, multiplying input tokens. Without a daily view, that spike is invisible; with it, you can trace the spike to a specific project and decide whether the pattern is worth changing.

Daily vs monthly vs session breakdowns

whoburnedmore surfaces three granularities. Day-level catches spikes. Month-level answers “is my plan paying off?” Session-level (available with npx whoburnedmore --sessions) shows individual conversation costs, which is the most useful view when you are debugging a single expensive workflow.

How is the estimated Claude Code cost calculated?

Claude prices input and output tokens separately, and the ratio matters a lot: output tokens typically cost 3–5x more than input tokens. whoburnedmore applies the public per-token price for each model you used and sums them:

For example, claude-sonnet-4-5 is priced at $3 per million input tokens and $15 per million output tokens as of mid-2026. A day where you sent 2M input tokens and received 180K output tokens would cost roughly $6.00 + $2.70 = $8.70 in API terms. That math is done automatically for each model in your log.

Where does Claude Code store its logs?

Claude Code writes JSONL files to a per-project directory structure under ~/.claude/projects/. Each session gets its own file, named by session ID. The JSONL format means every line is a valid JSON object representing one turn — with fields for the role, model name, and token counts. You can inspect these directly:

$ ls ~/.claude/projects/-Users-ab-code-myproject/  -Users-ab-code-otherproject/ $ ls ~/.claude/projects/-Users-ab-code-myproject/20260614T091234_abc123.jsonl   20260613T154500_def456.jsonl $ head -1 ~/.claude/projects/-Users-ab-code-myproject/20260614T091234_abc123.jsonl | python3 -m json.tool | grep -E 'inputTokens|outputTokens|model'    "model": "claude-sonnet-4-5",    "inputTokens": 48920,    "outputTokens": 3812,

This is the raw data whoburnedmore reads. The log path can differ: on some Linux setups Claude Code respects $XDG_CONFIG_HOME and writes to ~/.config/claude instead. whoburnedmore checks both locations automatically.

Does reading my Claude logs send anything to a server?

No code, no prompts, and no file names ever leave your machine. whoburnedmore reads the numeric fields — token counts, model names, timestamps — and sends only the aggregated totals to the leaderboard if you choose to submit. You can verify this with npx whoburnedmore --dry-run, which prints exactly what would be submitted. Or use npx whoburnedmore --local to open a local dashboard that stays entirely on your computer. 🛡️