Study finds maximum reasoning effort in AI models can cost far more and sometimes solve fewer tasks

A paper posted to arXiv in October 2026 examined how the reasoning-effort setting in frontier AI models affects cost and performance across eight models and four benchmarks. On tasks that every effort level solved, the highest setting cost up to 42 times more than the lowest. In some cases, higher effort also caused models to fail tasks that lower effort had completed, such as GPT-5.6-sol's drop from 76 to 73 solved Terminal-Bench tasks. The researchers found that starting at low effort and escalating only after a verified failure outperformed fixed maximum effort on both success rate and cost. The study was authored by researchers from Johns Hopkins, Together AI and Cornell.