Effort (reasoning effort)

Effort is the main dial for how hard a current model works on a request. On Claude it is set through the output configuration with levels from low through medium, high, xhigh, and max. Higher effort generally means more thinking, more thorough answers, more output tokens, and more time; lower effort means faster, cheaper, more direct responses. Defaults differ by model, so set it explicitly per feature rather than inheriting whatever the default is.

Effort has largely replaced the older knobs builders used to reach for. Instead of fixed thinking budgets or sampling tweaks, you pick an effort level per route: low for chat replies, classification, and anything latency-sensitive; higher for difficult analysis, multi-step decisions, and code.

The mistake is choosing by instinct. Run the same eval set at two or three effort levels and compare quality, latency, and output tokens. Often a lower level matches a higher one on easy features, which is free savings, and sometimes a higher effort on a small model beats a bigger model at low effort for less money.

Effort also pairs with routing: you can raise it only for requests a classifier marks as hard, keeping the common case fast.