“For most tasks in the Time Horizons benchmark, models exhibit deterministic behavior, either succeeding on every attempt or failing on every attempt.”