Reasoning effort level.

Reasoning effort level is a dial for how hard a model thinks before it answers. Some models are reasoning-sensitive, so this setting moves the quality a lot.

Turn it up (high, xhigh, max) and you get noticeably better work on hard problems. The catch is it burns Tokens faster and runs slower, so you match the level to the job. Quick task, keep it low. Gnarly coding or a tricky plan, turn it up.

Remy's example shows the gap: "the coding score drops from 63 to 42 just going from xhigh to high." Same model, same prompt, only the effort changed. So if the output feels lazy on something genuinely hard, check this before you blame the model.