Statistical audit of METR time-horizon plot: a 10x jump isn't always 10x

Two Berkeley statisticians (Nguyen and Fithian, October 2026) re-examined METR's time-horizon plot using 228 tasks, 79 task families and 26 models. They don't dispute the exponential trend, but find the human-time-to-AI-difficulty link is nearly flat between 2 and 30 minutes, so a jump from 3 to 30 minutes is far easier than one from 30 minutes to 5 hours.

• Data: 228 tasks, 79 families, 26 models, 5-fold cross-validation by family
• Conversion curve is nearly flat from 2 to 30 minutes of human time
• 83% of model runs on a task are unanimous vs 60% expected
• About 29% of tasks rely on expert time estimates, not timed runs