Effort level: low

Where this tier actually belongs

Low effort is the right default for tasks where more reasoning genuinely wouldn't change the outcome — a classification with a small, well-defined set of possible answers, an extraction task against clearly structured input, a subagent handed a narrow, well-scoped lookup. The common thread across good uses of this tier isn't "the task looks easy" — it's that the task has little genuine ambiguity for reasoning to resolve in the first place.

The failure mode is quiet, not loud

Applied to a task that actually needed more reasoning, this tier doesn't error or produce an obviously broken response — it under-thinks, and the result can look complete while missing something a closer read would have caught. That makes misapplying this tier riskier in a subtle way than picking a level that's too high: an overspend is visible on the bill; an underspend that produced a subtly wrong answer often isn't visible anywhere until it causes a downstream problem.

Raising effort, not adding instructions, is the fix

When output at this tier is consistently missing something, the fix is moving to a higher effort level, not layering more instructions onto the prompt asking the model to "think harder" or "be more careful" — that kind of prompt addition doesn't reliably substitute for actual reasoning depth the task needed and this tier wasn't built to provide.

Related

See the effort & thinking cost estimator for where this tier sits on the cost curve relative to the rest of the range, and effort level: medium for the next step up.