Your AI Coding Bill Isn't Paying for Code

Where the money actually goes when you use Claude Code, Cursor, or Codex — and the three leaks you can fix today.

---

Every month, developers open their AI bill and feel the same confusion: "I didn't even code that much." The number is always bigger than expected, and the dashboard never explains why.

I've spent the last few months digging into how AI coding tools actually consume tokens. Here's the uncomfortable truth: most of your bill isn't paying for code. It's paying for waste. Three leaks account for the majority of it.

Leak 1: The idle-session cache tax

Modern coding assistants cache your conversation context so follow-up prompts are cheap. That's great — until you walk away.

When you return to an idle session hours later, that cache has expired. The entire context gets re-read at full price. One developer found that 22% of a 33-day Claude Code spend came from nothing but returning to idle sessions. Not new work. Not better answers. Just re-reading old context at full price.

The fix: start fresh sessions for new tasks instead of resurrecting dead ones. If your tool shows cache read vs. fresh input tokens, watch that ratio — it's the single most honest number on your bill.

Leak 2: Subagents that bill like executives and produce like interns

Subagents are powerful — a research agent here, a test-writing agent there. But each one carries your full context with it, and they don't share nicely.

Another developer measured it: subagents were 48% of the bill while producing 0.9% of the output. Nearly half the money, less than one percent of the value. The convenience feels free because you never see a price tag per agent spawn.

The fix: be deliberate about spawning subagents. Ask yourself: does this task actually need isolated context, or can the main agent handle it? For routine tasks, one agent with a good prompt beats five agents with duplicated context.

Leak 3: The wrong model for the job

Not every task needs the flagship model. Drafting a README, renaming variables, writing boilerplate tests — these burn premium tokens for commodity work.

Most tools now let you pick models per task or per agent. The fix: route simple work to smaller, cheaper models and save the heavy model for architecture decisions and hard debugging. This alone can cut 30–40% off a typical bill with zero quality loss on the easy stuff.

How to actually see your waste

Here's the meta-problem: no vendor shows you the full picture. Each tool reports its own slice, in its own format, and none of them will tell you "hey, 48% of your spend produced nothing."

What works:

1. Log every run in one place — model, input tokens, output tokens, cache hits, cost. A simple spreadsheet works.
2. Grade your sessions — which tasks were worth it? You'll quickly see patterns (Friday-afternoon "quick experiments" are usually the worst offenders).
3. Set a budget with alerts — decide what AI coding should cost you per month, and get warned at 50%, 80%, 100%. Without a number, every bill feels "about right" until it isn't.
4. Review weekly — 10 minutes every Friday looking at where tokens went will save you more than any prompt trick.

The bottom line

AI coding tools are genuinely worth paying for — the productivity gain is real. But right now most developers are paying a 30–50% "waste tax" on top of the real price, simply because they can't see where tokens go.

Start with the three leaks above. Measure for one week. You'll probably find your money.

---

I write about AI tooling costs and practical ways to keep them under control. If this helped, follow along — more teardowns coming.