Clarify whether GPT-5-mini’s reasoning.effort directly controls CoT length

Ascertain whether the reasoning.effort parameter in GPT-5-mini directly controls the number of chain-of-thought tokens generated (i.e., CoT length), as opposed to modulating some other aspect of inference-time reasoning.

Background

In the paper’s evaluation on GPT-5-mini, the authors compare attack success rates under different reasoning-effort settings (minimal, low, high) and observe that success peaks under low effort, suggesting that reasoning-effort and CoT length may be related but distinct controls.

Immediately following this result, the authors note uncertainty regarding the official description of the parameter, explicitly stating they are not sure whether reasoning.effort constitutes a direct control of chain-of-thought length, motivating a need to determine the parameter’s precise operational effect.