“Early data suggests OpenAI's o3 model will score under 30% on the upcoming ARC-AGI-2 benchmark, even with high compute.”