Almost every deployed model misreports its own progress mid-execution.

Evaluating task-progress reporting on τ²-bench and StageIF, reliability depends on which stage the task has reached: most deployed models lose accuracy once work is under way and recover once done, while the newest generation closes the mid-task dip and instead under-reports completion at the finish line. arXiv 2609.08589 The paper's explicit conclusion is to stop gating control flow on the model's own state reports.