Found first: a primary source the press has not covered yet.
A paper submitted to arXiv on October 25, 2026 finds that on tasks every effort level can solve, the highest reasoning effort costs up to 42x as much as the lowest, and higher effort frequently fails on tasks that lower effort solved. "What Effort Level Should I Choose? On the Economics of Reasoning Effort in LLM Agents" goes on to show that a simple escalation strategy outperforms fixed maximum effort across all models and benchmarks tested, at lower cost.
What the source says
Authors Kaiser Sun, Yongchan Kwon, Zain Hasan, Shang Zhu, Kaitlyn Zhou, and Federico Bianchi (no institutional affiliations are listed on the arXiv page) studied how selecting inference-time compute levels affects cost and task outcomes across proprietary models and benchmarks. On tasks solvable at every effort level, the highest effort level costs up to 42x the lowest. Higher effort changes agent behavior: agents verify their own work more heavily and search for solutions more often rather than solving directly, without improving outcomes. A strategy of starting at the lowest effort and escalating only after a verified failure matches or improves on fixed top effort across all models and benchmarks, at lower cost.
Why it matters
The authors frame reasoning effort as a "credence good," a service where buyers cannot fully assess whether the purchased computation was necessary, even after observing outcomes. That framing matters because it means practitioners have no reliable feedback loop to calibrate their choices. Anyone calling a frontier model API with reasoning effort set to maximum is likely overpaying on a significant share of tasks. The escalation heuristic the paper describes is vendor-agnostic and requires no task-specific tuning.