ClockBench evaluates whether models can read analog clocks - a task that is trivial for humans, but current frontier models struggle with.