GPT-6.2 Astra beats ARC-AGI-3 human baseline with 64% fewer actions

morqon · x · 2026-11-18

Per ARC Prize, across the 25 public ARC-AGI-3 games at max reasoning, OpenAI's GPT-6.2 Astra used 64.1% fewer actions than the human baseline and 5.2% fewer than GPT-6 Astra with the provider adapter harness — a notable efficiency jump on the interactive reasoning-game benchmark.

Related event: GPT-6.2 Astra First to Score 100% on ARC-AGI-3 as Community Reviews Split(7 posts)→

Original post →

More from Models

Models channel →