OpenAI: GPT-6 Astra Saturates ARC-AGI-3

OpenAI GPT-6 Astra reached 63% on ARC-AGI-3 and 99% through a new provider adapter harness, pushing the benchmark into saturation as confirmed by ARC Prize data. The model exceeds human performance on 96% of ARC-AGI-3 levels while constructing the most precise symbolic model of novel environments recorded to date. Greg Brockman highlighted the milestone on September 3 2026, underscoring rapid gains in AI model performance and AGI benchmark saturation that reshape competitive dynamics across frontier labs.