Cut Your AI Token Bill With a 20-Minute Audit

Your agent can reach your email, your files, and your calendar. Twenty minutes and a microphone to decide what it's actually allowed to do with them.

I found out I was an AI over-spender the same way I found out I was a cloud over-spender in the 2010s — the bill told me. My Manus account runs out of credits most months before the billing cycle ends, and when I finally ran the numbers on my Claude Code usage, the pattern was embarrassing: I was paying for the same conversation history over and over, letting agents read half a repo to fix one file, and running frontier models on work an entry-level model handles fine.

So here's the thing nobody tells you about token waste: it isn't a leak, it's a subscription. An agent has no memory between turns — every turn it re-sends the entire conversation back to the model, and you're billed for all of it again. A token you wasted on turn 3 keeps costing you on turns 4, 5, and 6. The good news: the fix takes about twenty minutes, and most of it is habit, not tooling.

// The Takeaway: Run ccusage to get a baseline of exactly where your Claude tokens go, then attack the four sources of waste: what goes into the context window, what tools dump back, what the model writes, and what gets re-billed every turn. Same monthly spend, sharper output — you're not using AI less, you're making every token earn its place.

Run npx ccusage@latest in your terminal now → Twenty minutes. Free, open source. No signup, no config.

Every AI conversation runs on tokens — the smallest unit of data a model processes, roughly four characters of text. Your context window is the model's working memory, and it's the most valuable real estate you have. Fill it with the three files that matter and you get sharp work; fill it with forty files of noise and you get a distracted intern — one that charges by the word.

// The real shift: Stop thinking of your AI subscription as all-you-can-eat and start thinking of the context window as a budget you spend every single turn. The skill isn't spending less — it's knowing what each token is buying you. That's the same mental move cloud-native companies made with FinOps a decade ago, just closer to the keyboard.

Every wasted token falls into one of four buckets:

No terminal required. Set up Projects to load your reference files once instead of pasting them into every conversation, and turn on memory so you stop re-explaining your role, your company, and your preferences — that's hundreds of tokens of setup you're currently paying to repeat in every chat. Batching helps too: one structured prompt with three asks beats three separate prompts that each drag the full history along behind them.

The market just made conservation a business skill. Uber burned through its entire 2026 Claude Code budget in four months, one enterprise reportedly spent half a billion dollars on Claude in a single month, and Gartner projects AI coding costs will rival developer payroll by 2028. The habits in this Lesson are the individual version of the governance your company is about to mandate anyway. Learn them now and you're the person in the room who already knows where the tokens go — Friday's Deep Dive picks up the story at the enterprise level.