Claude Opus 5’s “max effort” setting inflates the price of a routine request from $0.76 to $19.21 while delivering essentially the same functional output. The extra spend buys an internal audit pass, not a better solution, and only shows measurable gains on tasks that start with low test coverage.
What the test showed
The experiment compared Claude Opus 5’s default effort level with the “max effort” knob on two prompt types: everyday coding chores and deliberately hard problems.
- For a typical task, the low-effort run finished in two minutes and cost $0.76. Cranking the setting to max pushed the bill to $19.21, yet the requirement-coverage score— the metric the model reports for how well it met the brief— stayed identical.
- The transcript shows a shift from creation to revision. The model stopped generating fresh code and started polishing what it had already written. Edits outnumbered new writes by 2.4-to-1. Calls to the “read” tool rose eighteenfold, and invocations of the “bash” tool climbed sixfold. In practice the model re-read modules, re-ran its own tests, performed linting and even mutation testing without being asked.
The “max effort” knob does not introduce a new algorithm; it simply raises the budget the model can spend. Once the budget is large enough, the model switches to a self-audit mode, looking for any tweak it can justify spending on.
Why the cost spikes
When the model decides to audit, every extra read or bash call adds to the bill, and the multiplier effects quickly balloon the total cost.
The audit mode is an explicit design choice. The model treats the larger budget as permission to “look for something worth fixing.” If nothing appears improvable, the extra spend yields no functional benefit.
When higher effort makes sense
The audit mode shines only when the initial output leaves room for improvement. In a Go project with test coverage of 0.73, pushing the effort to max lifted coverage to 0.88.
Conversely, a Python task that already achieved 0.98 coverage saw no change when the budget was increased. The model simply re-checked the same high-quality code, inflating the cost without adding value.
Potential downsides
- Budget blowout – Users accustomed to the low-effort price point may be surprised by a twenty-five-fold increase for the same deliverable.
Practical guidance for developers
- Keep routine prompts at the default effort level. You get the same functional result for a fraction of the price.
- Reserve “max effort” for code that fails to meet a clear quality threshold— low test coverage, missing lint warnings, or other measurable gaps.
- Treat the setting as a separate mode: an optional self-review pass rather than a speed dial for better answers.
Takeaway
Claude Opus 5’s max-effort switch trades money for an internal quality check, not for better code. Use it sparingly, only when your baseline results leave a quantifiable hole to fill; otherwise, the cheap default delivers the same outcome without the audit-mode price tag.
Community discussion: https://t.me/GyaanSetuAi