"Humanity's last exam, this currently looks like it's state-of-the-art, just three points behind GPT 5.4 Pro, and it's actually currently better than other models when it doesn't use tools."