I burned a week of Claude quota in 3 days. Then I found the setting eating it.
Описание и применение
Автор потратил неделю квоты Claude Ultra за три дня, но после перехода на более экономичные настройки и использования навыка Rightsize сократил расходы до 0,02 $ за запрос, сохранив точность.
Задача
The author spent a full week of Claude Ultra quota in only three days, incurring high costs without clear justification.
Результат
Using lower effort settings and the Rightsize skill cut costs from $2.03 to $0.02 per run (up to 99% savings) while preserving accuracy on checked tasks, and reduced test run time from 11.4 s to 0.7 s on a 40‑module repo.
Рабочий процесс
- Benchmark Claude Ultra by running 100 scripted tests and grading results.
- Compare cost and effort levels (low, medium, high, max) for Opus and Haiku models.
- Run Haiku at lowest effort on a regex‑engine task and validate against Python’s re on 3,000 patterns.
- Apply low‑effort models to everyday work (search, rename, parser fixes) and measure cost and pass rate.
- Implement Rightsize skill to auto‑select model/effort and run only affected tests.
- Audit local transcripts to track token usage without sending data outside the machine.
Доказательства
I burned a full week of my Claude Ultra quota in 3 days.
at the lowest effort: $0.19 at the highest effort: $2.03
It matched all of them. Three runs out of three. For $0.02.
$0.352 became $0.041. 88% less, and the same 6 passes out of 6.
It picks the model and effort for every step, and saves the heavy thinking for the work that truly needs it.
- Дата записи: 2026-11-03T21:14:03.514Z