ZeroBench
A set of 100 image analysis puzzles built to be hard enough that even the best models mostly fail them.
ZeroBench takes the opposite approach to most visual tests. Instead of collecting a large pile of ordinary images, it uses just 100 questions, each deliberately constructed so that models cannot get there by pattern matching. Counting things arranged awkwardly, reading a chart where the interesting detail is small, reasoning across several parts of one picture at once.
The name comes from the ambition: when it was published, the goal was a benchmark on which frontier models score zero. They no longer do, but scores stay low enough that a difference of one or two points between two models is well inside the noise. Useful for tracking progress over a year, much less useful for declaring a winner this week.