4 ms·

ARC-AGI-3 is interesting, but it’s not necessarily aligned with what models are being optimized for right now (reasoning + usefulness in real tasks).