The Innermost Loop

The cost of reasoning has officially decoupled from the cost of human labor. OpenAI has released GPT-5.2 Thinking, the first model to beat human experts on GDPval, a rigorous benchmark testing well-specified knowledge work across 44 distinct occupations. By outperforming professionals on 70% of these tasks at less than 1% of the cost, the model signals the end of the premium on cognitive drudgery. It also cracks the 90% threshold on ARC-AGI-1 for just $11.64 per task, a 390x deflation in the…