Maximum Effort By Default
If the model has a hardest setting, the temptation is to leave it there. It costs more and reasons worse.
Reasoning effort is a dial, and leaving it at maximum feels like buying quality. It is not free: cost rises, latency rises, and on simple tasks the extra reasoning produces overthinking rather than accuracy. The working default is one notch below maximum, reserving the top for irreversible decisions.
Pattern
The model exposes a reasoning-effort setting with 5 positions. Higher is better, the highest is best, and nobody wants to be the person who chose less thinking. So it stays pinned at maximum.
Why it looks right
Quality is the thing everyone is optimising and effort is the visible lever for it. Turning it down feels like accepting worse output to save money, which is a bad trade on anything that matters.
There is also no feedback. A task done at maximum effort produces no evidence that a lower setting would have produced the same answer.
Why it fails
Three costs, and the third is the one that surprises people.
Money and latency scale with it, on 100% of tasks, including classification and lookups where the answer was never in doubt.
The budget is finite, so effort spent on trivia is unavailable for the decision at the end of the day that actually needed it.
More reasoning is not monotonically better. On simple, well-specified tasks the extra passes produce second-guessing — an answer that was correct in one step gets revised into something more elaborate and less right. Overthinking is a real failure mode, not a rhetorical one.
Instead
Match the setting to the task, with a written default so the choice is not made fresh each time:
triage, classification, lookups → low · everyday work → high · deep multi-step agentic runs → extra · maximum: irreversible decisions only
Two rules keep it honest. The workhorse is one notch below maximum, not maximum — reserving the top makes it meaningful. And when the operator signals urgency — quickly, briefly — that is a reason not to raise the setting, since the request was for a fast answer and a slow one fails it regardless of quality.