OpenAI's GPT-6 Astra on ARC-AGI-3

GPT-6 Astra achieved 62.7% accuracy on ARC-AGI-3’s semi-private test for $26 K and 99.9% accuracy for $19 K using a provider-adapter harness that preserves reasoning state and compacts longer conversations. It outperformed the median human on 96% of levels, using fewer actions and converting unfamiliar game mechanics into compact symbolic models with a self-generated domain-specific language.