GPT-5.2 Pro Scores Higher on ARC-AGI Than o3 Did — at 400x Lower Cost

OpenAI's cost per task on the ARC-AGI benchmark dropped from $4,500 to under $12 in roughly a year, while accuracy actually improved — a stunning efficiency gain that reshapes the economics of frontier reasoning.

The cost of frontier AI reasoning on the ARC-AGI benchmark has collapsed by nearly 400x in a single year, as @rohanpaul_ai highlighted on Monday. When OpenAI's o3 model was evaluated on ARC-AGI roughly a year ago, it scored 88% accuracy at a cost of approximately $4,500 per task. GPT-5.2 Pro now scores 90.5% — a modest but real accuracy improvement — at just $11.64 per task.

Unlock the full briefing

Get every story in today's briefing, the full archive, and the daily AI intelligence brief.

All stories today

Full archive

Daily brief

Cancel anytime. Payments powered by Stripe.