Quick answer
There is no normal number. A light day of chat-style help can be thousands of tokens, while a day of agent runs in Claude Code or Codex can reach tens of millions, mostly cache reads. Measure your own days and compare against your baseline.
Why the range is so wide
- Agents reread context on every turn, so long sessions grow fast.
- Large files and logs add input in big chunks.
- Reasoning models add output tokens while they think.
- Cache reads can make up most of an agent session.
Measure your own baseline
UsageAtlas records tokens per day across Claude Code, Codex, Cursor, OpenCode, and other local sources. Today compares the current day with your recent baseline. Insights show which hours and weekdays run hottest.
After two weeks you know your normal. After a month you can tell whether a plan fits. Download UsageAtlas to start.
Tokens are not your limit
Subscription limits for Claude Code and Codex are percentages of a hidden allowance, not token counts. Watch both. Tokens tell you what happened, the window tells you what is left. Token types explained.
Questions, answered
Is 10 million tokens a day a lot?
For agent work it is common, because cache reads inflate totals. Check the split between new input, output, and cache before judging.
How do I see my tokens per day in Claude Code?
Claude Code shows the current session. UsageAtlas keeps a daily total from local history, so you can open any past day.
Does UsageAtlas count tokens from every tool?
It counts tokens from tools with local history, such as Claude Code, Codex, Cursor, and OpenCode. Credit-only connections show balances instead.