I burned a week of Claude quota in 3 days. Then I found the setting eating it.

Description and use

The author spent a week’s Claude Ultra quota in three days, but by switching to lower effort settings and using the Rightsize skill, reduced costs to $0.02 per request while maintaining accuracy.

Problem

The author spent a full week of Claude Ultra quota in only three days, incurring high costs without clear justification.

Outcome

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.

Workflow

• 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.

Evidence

I burned a full week of my Claude Ultra quota in 3 days.

Source

at the lowest effort: $0.19 at the highest effort: $2.03

Source

It matched all of them. Three runs out of three. For $0.02.

Source

$0.352 became $0.041. 88% less, and the same 6 passes out of 6.

Source

It picks the model and effort for every step, and saves the heavy thinking for the work that truly needs it.

Source