AI Usage Audit — report
Period: last 14 days · Total: 236M tokens
The headline: one session is your entire problem.
Your top session burned 128M tokens — 54% of everything you used in two weeks. It was a marathon, continued from a summary. That restart helped once, then the session grew right back past 100M. This single session, not your daily work, is what pushes you into the limit.
Why it happened (and why it's not your prompts).
Your actual typed input over 14 days was ~75K tokens. That's 0.03% of the total. The other 99.97% is the model re-reading its own context on every message. You are not "using too much" — one long session is dragging its whole history behind it, and you pay to re-read it every turn.
Your cache is fine — ignore anyone who says otherwise.
Cache hit ratio: 99.97%. Caching is working perfectly. Your leak is NOT a caching problem. Advice like "improve your caching" is wrong for your usage. Your leak is 100% session length.
The 3 fixes, in order of impact:
- 1.
Close sessions at task boundaries (biggest win). Compacting won't save you — it empties the trailer but the session refills it. What cuts usage is closing the session. When a task ends, start fresh. Need continuity? Paste a short summary into a new session instead of continuing the giant one. If that 128M session had been capped around 30M and split, you'd have avoided ~90M tokens — roughly 38% of your entire two-week usage.
- 2.
Route simple work off Opus. 46% of your tokens ran on Opus. Refactors, boilerplate, and simple edits would look identical on a cheaper model. Keep Opus for genuinely hard reasoning, send the rest down. Even moving a third of that off Opus frees ~15M tokens of headroom.
- 3.
Stop treating long sessions as productive. They feel efficient but eat your limit like a second subscription. Your output was 0.4% of everything — you're not paying for what Claude writes, you're paying for what it re-reads. Shorter, scoped sessions are the whole game.
Bottom line.
Between session hygiene and model routing, cutting 30-45% of your usage is realistic. On the $100 plan that means you stop hitting the weekly wall — without paying $200.
