GPT-6 Astra Saturates ARC-AGI-3 Benchmark with Record-Breaking Performance
OpenAI's latest AI model, GPT-6 Astra, has reached a significant milestone in its development. According to data from ARC Prize, the model has saturated the ARC-AGI-3 benchmark, achieving a score of 63% on this metric and 99% through a new provider adapter harness.
This achievement is notable because GPT-6 Astra has exceeded human performance on 96% of the ARC-AGI-3 levels. Additionally, it has constructed the most precise symbolic model of novel environments ever recorded to date.
Greg Brockman highlighted this milestone on September 3, 2026, emphasizing the rapid gains in AI model performance and AGI benchmark saturation that are reshaping competitive dynamics across frontier labs.