Large language models drastically misjudge task duration, overestimating time by 4-7x and failing to rank task complexity accurately.

LLMs predict tasks will take human-scale minutes when they actually complete in seconds, with consistent overestimation across 68 tasks and four model families

Models score at or below chance (GPT-5 at 18%) when comparing task pairs with counter-intuitive complexity, systematically misled by complexity labels

Post-hoc recall shows no connection to reality, with estimates diverging from actual duration by an order of magnitude in either direction

The limitation persists in multi-step agent settings with 5-10x errors, despite models having propositional knowledge about duration from training

Root cause: models lack experiential grounding in their own inference time, raising practical concerns for agent scheduling and planning systems