← cc-safe-setup Field notes · 800-hour autonomous run · independent

How I made Claude Code cheaper — the 5 changes that worked

Over an 800-hour autonomous run I tried a lot of things to bring the token bill down. Most of the gains came from just five small habits. Here they are — each one is a win you can copy today, and every one is free.

The good news up front: you don't need a complicated setup to spend a lot less. The five changes below are the ones that kept paying off, turn after turn — leaner sessions, calmer bills, and faster replies as a bonus. Start with the first one; it did the most.

The five wins

1. A fresh session per task — the single best change

Starting a new session (/clear) whenever I switched topics turned out to be the biggest lever by far. A long session keeps re-sending its whole history every turn, so the per-turn cost quietly climbs as it grows. Clearing between unrelated tasks keeps every session short and the prompt small.

Result: each turn stays cheap instead of getting ~10× pricier by turn 200, and replies come back faster because the model re-reads less.

Habit: one task, one session. When you move to something unrelated, /clear and begin fresh.

2. Sonnet for the everyday work — the biggest line-item drop

Opus is wonderful for hard reasoning, but most routine work — edits, refactors, file shuffling, running commands — lands just as well on Sonnet, which is several times cheaper per token. Setting Sonnet as the default for ordinary turns and reaching for Opus only when a problem actually needs it was the change that moved the monthly number the most.

/model sonnet     # everyday work
/model opus       # bring it back for the genuinely hard problems

Result: the bulk of my turns ran on the cheaper model, and I honestly couldn't tell the difference on routine work.

3. Compacting on purpose kept long sessions smooth

When a session does need to run long, calling /compact yourself at a natural breakpoint (after finishing a sub-task) keeps it tidy and predictable. Doing it on purpose, rather than waiting for the automatic trigger to fire deep in a huge context, kept the summaries small and the flow uninterrupted.

Result: long sessions stayed lean and steady, with no surprise pauses to re-summarize a giant transcript.

4. Watching where the tokens go made costs calm and predictable

Once I could actually see the numbers, the whole thing stopped feeling mysterious. The free, open-source ccusage reads your local session logs and shows a per-session breakdown, so you learn your own normal range and spot anything unusual early — no bill-day surprises.

Result: spending became something I could plan around instead of guess at, which made every other habit on this list easy to keep.

Tip: a one-line check after a heavy session is enough to keep your sense of "normal" current.

5. Keeping context lean made every turn faster and cheaper

Loading only what a task actually needs — trimming the files, tools, and MCP servers that get re-sent every single turn — pays off continuously, because anything in the context is paid for on every turn it survives. A tidy context is cheaper and snappier.

Result: lighter prompts, quicker replies, and a smaller bill, all from the same habit.

The shortlist

If you do just these, you'll feel it on the next bill:

Want the longer version? The free field reference of cost patterns lists every common cause with the fix for each, and the cc-safe-setup hooks watch your own usage for free so good habits are easy to keep. If you want the full worked-through playbook — per-task model routing, cache hygiene, and a token-reduction checklist with the math — the Token & Cost Survival Guide ($19) collects everything that worked in one place.
cc-safe-setup · GitHub · Field notes from an independent 800-hour autonomous run; your mileage will vary with your workload. Model and pricing details change — confirm current rates against Anthropic's site.