Low, Medium, High: A Practical Guide to Opus 5's Effort Toggle
Opus 5's per-request effort toggle is genuinely useful, but only if you actually vary it — the easy failure mode is defaulting every request to "high" out of habit, which quietly erases the cost savings the toggle was built to offer. A rough starting heuristic:
- Low: classification, extraction, simple formatting, anything where a clearly-defined correct answer exists and the task doesn't require multi-step reasoning.
- Medium: most everyday coding and writing tasks — the default worth reaching for unless you have a specific reason to go higher or lower.
- High: genuinely hard, ambiguous, or high-stakes requests — an architecture decision, a subtle bug, anything where getting it wrong is expensive to discover later.
Measure before you tune
Before hand-tuning effort levels across your request types, instrument your calls to log effort level alongside outcome quality (however you currently measure it — user satisfaction, error rate, review pass rate). A few weeks of that data will tell you far more about where "medium" is actually sufficient than intuition will.
⭐⭐⭐
anthropic.com