Why the bill moved

Costs rarely rise because you did one dramatic thing. They rise because of five ordinary ones, all of which look free at the moment you do them.

1. The conversation that never ended

The biggest and least visible.

L1 established that the whole history is re-sent every turn. A working session that runs all afternoon is therefore paying for its own transcript, repeatedly, and the cost per message climbs steadily even though your questions stay the same size.

Fix: start a new conversation when the topic changes. Not for tidiness — because you stop paying for two hours of unrelated history.

This is also P1·L5’s instruction drift arriving as a bill. Long threads cost more and follow instructions worse, which makes the fix free of trade-offs.

2. The document you attached once

Attach a 60-page PDF and it enters the context. Ask nine follow-up questions and that PDF is sent ten times.

Fix: E5 and P1·L6’s advice, now with a price attached. Extract the section you need and work with that. It is cheaper, and L6 showed it produces better answers — one of the rare changes that improves both.

3. Retries you did not count

A run that fails still costs. A malformed answer you regenerate costs twice. A prompt that needs four attempts costs four times, and the three you discarded were billed exactly like the one you kept.

Fix: this is what P1·L7’s evaluation set is for. A prompt that works first time is a cost improvement as well as a quality one, and vague prompts are expensive in a way that is invisible until you look.

4. The automation nobody watched

The one that produces a genuinely alarming invoice.

An automation that runs per email, per row or per webhook has no natural limit. A loop that retries on failure can retry indefinitely. A trigger that fires on its own output can run away entirely.

Fix: caps and alerts before it runs unattended. L7 covers this properly — it is the single most important lesson for anyone building.

5. The biggest model, by default

Using a capable model for everything is the quiet, steady version of this. It does not spike; it just costs several times what the work needed.

Fix: L3.

What to actually do first

In order of saving per minute spent:

1. Start new conversations. Free, immediate, and helps quality.
2. Trim what you attach. Cheaper and better answers.
3. Ask for shorter output where you do not need length — output tokens cost more.
4. Check for caps on anything automated.
5. Right-size the model (L3).

None of these involves doing less work. That is the point: the common cost problems are waste rather than usage.

Where the meter is not running

If you are on a flat subscription, none of this appears on a bill — it appears as hitting your limit sooner. The fixes are identical; the symptom differs.

Try it now (5 minutes)

Open your longest current conversation. Ask yourself whether the first half is still relevant.

If not, start a fresh one and paste only what matters forward. You have just made every subsequent message cheaper and more likely to follow instructions.

Check your understanding

1. The most common cause of quietly rising cost:

Using a more capable model Long conversations, since the whole history is re-sent each turn Asking more questions per day

2. A prompt that needs four attempts costs:

The same as one, since only the final answer counts Four times, because discarded attempts are billed identically Less, because failed runs are not charged

3. Which change improves both cost and answer quality?

Attaching only the section the task needs Using a larger model Asking for longer answers

Check answers

Recap

Five causes: threads that never end, documents re-sent with every follow-up, uncounted retries, unwatched automations, and defaulting to the biggest model. The first two are free to fix and improve quality at the same time. None of the fixes involves doing less work, because the usual cost problem is waste rather than usage.