The ceiling on abstract reasoning tasks is fracturing as frontier models like GPT-5.2 paired with agentic scaffolds routinely exceed prior state-of-the-art by double-digit margins on holdout evaluations like ARC-AGI.
Poetiq achieved 75% on ARC-AGI-2 public-eval using GPT-5.2 x-high without model-specific training or optimization, slashing costs to...