Study: Max Reasoning Effort Costs Up to 42x More and Sometimes Solves Less

An arXiv paper (2610.34491) testing 8 models on 4 benchmarks found that max reasoning effort costs up to 42x the lowest level on already-solved tasks and can reduce the number of solved tasks. Starting low and escalating only after a verified failure beats fixed max effort on both success and cost.

• On Terminal-Bench, low → xhigh raises cost 2.51x for GPT-5.6-sol and 3.04x for Claude Sonnet 5
• On SWE-Bench Pro GPT-5.6-sol goes from $0.27 to $1.24 per task (4.54x)
• GPT-5.6-sol at high → xhigh loses 8 tasks while gaining 5: 76 → 73 of 89
• Laddering from low and escalating after failure beats fixed max on success and cost