Claude Code effort levels: what one real-world test found
A creator tested Claude Code’s effort settings on Opus 5.5 instead of relying on the labels. The main takeaway: medium was the best default in this test; higher effort often cost much more time without improving the result.
1. What he tested
- Model: Claude Opus 5.5.
- Effort levels: low, medium, high, X high, and max.
- Method: five runs at each level on the same task, checked by a hidden test suite running in real Chrome.
- Task: write a filter that removes every way JavaScript can run from a web page.
- Test suite: 90 checks.
2. The headline results
~2½ minutes
~9 minutes
~17 minutes
X high and max
X high: none of five runs finished within the 35-minute cutoff. Max: one run finished, but it missed five checks.
“Stay on medium, move to high when a bug is hiding, and skip X High and Max unless you can wait.”
3. Other experiments he describes
Bug hunt
With five planted bugs, every level fixed all five. Low took about one minute; max took 13 minutes and cost almost nine times more.
CommonMark test
Using the full CommonMark specification—652 official tests—Opus on low passed all 652 twice. Sonnet on low reached 83% in one run. His conclusion: for this task, the model mattered more than the effort level.
Front-end design
For a landing page, low produced a polished first screen in about 100 seconds. X high took about 15 minutes and added more sections farther down the page.
4. His explanation of the settings
- X high: he says it sits between high and max; the “X” likely means “extra,” but the docs did not explain the name.
- Ultra code: he says this is not a separate effort level. It is a setting for running workflows on top of the selected level. In his test, the
ultra codeflag also set X high. - /effort auto: he says this does not choose the best level for each task. It clears a saved level and falls back to the model default; for Opus 5.5, Sonnet 5.5, and Haiku 5.5, he says that default is medium.
- Changing effort: about 99.8% of the cache was reused when he switched effort. Switching from Opus to Sonnet rebuilt 47,000 tokens because each model has its own cache.
5. Practical takeaway
- Use medium as the starting point for ordinary work.
- Move to high when a difficult bug or hidden issue needs more investigation.
- Use X high/max only when the task is unusually hard and waiting—and paying more—is acceptable.
- Do not assume auto means “smartest level for this task”; in his description, it effectively means the saved/default level.
6. Important caveat
These are the creator’s measurements from a few tasks, not an independent benchmark. The results may change with the codebase, prompt quality, model version, test design, tool access, and time limits. The strongest defensible conclusion is not that low or medium is always best—it is that more effort is not automatically better value.
7. Source
Original video:
https://youtube.com/shorts/DaYql4hCWSg
This page is a cleaned summary of the spoken content and preserves the creator’s claims as claims. It is not an official Anthropic benchmark or documentation.