Reasoning effort with Meridian

Control how much compute a model spends thinking before it answers. Higher effort = deeper chains of thought, better accuracy on hard problems, and higher latency.

Modelreasoning_effortNotes
gpt-5-codexlowmediumhighDefault: medium
gpt-5-1lowmediumhighDefault: medium
grok-4-20-reasoninglowmediumhighDefault: medium

Important

Set max_tokens ≥ 2048 or use reasoning_effort=low to avoid empty responses. Reasoning tokens consume the max_tokens budget — if the model spends everything on internal thought, the visible answer is truncated to zero.

Back to docs