Opus 4.8 posted a state-of-the-art 1.5% of human efficiency on ARC-AGI-3, tripling GPT-5.5’s score. One and a half percent of a human, and it is the best in the world — the benchmark that still humiliates every model is also the one where quarterly gains now come in multiples.
Elsewhere: hackers took over the Obama White House Instagram by politely asking a support bot to change the email.