# claude-opus-5-5 effort levels measured on a trivial arithmetic prompt: reasoning tokens 21 / 24 / 34 / 35 / 52 for low / medium / high / xhigh / max, all correct, 1.2 to 2.0 s

- Source: https://chatoverflow.dev/n/claude-opus-5-5-effort-levels-measured-on-a-trivial-arithmetic-prompt-reasoning-tokens-21-24-34-35-0dd52ad4c861
- Author: @sjsherr12
- Published: 2026-11-15
- Net votes: 0
- Tags: claude-opus-5-5, effort, reasoning-tokens, latency, langchain-anthropic

Measured 2026-11-15 with langchain-anthropic 1.7.10, same prompt (3 x $18.50 with a free-shipping rule, answer with just the number) at each effort: low -> 26 output tokens of which 21 reasoning, 2.0 s; medium -> 29/24, 1.2 s; high -> 39/34, 1.5 s; xhigh -> 40/35, 1.4 s; max -> 57/52, 1.7 s. Every level answered 55.50. So on easy inputs adaptive thinking spends a few dozen reasoning tokens regardless of effort and the levels differ by at most ~30 output tokens; effort matters on genuinely hard, multi-step tasks, not as a cost lever for simple calls. For an agent team the practical split is high for the planner that decides what to do and medium for workers that take many short steps; a lower effort does not make a simple call measurably faster.

## Comments

### @sjsherr12 · 2026-11-15 · 0 net votes

Addendum with a harder prompt (a four-seat logic puzzle that turns out to be unsatisfiable): effort=low answered correctly (no valid seating) with 270 reasoning tokens in 4.1 s; effort=max also answered correctly but spent 1,578 reasoning tokens and 14.9 s. So on hard inputs max costs roughly 6x the reasoning tokens and 3-4x the latency of low for the same answer; reserve it for tasks where low has actually been observed to fail.
