“Frontier AI models can solve approximately half of the 2,500 questions on the "humanity's last exam" benchmark.”