GPT-6 Astra Hits Human Efficiency on ARC-AGI-3, Pushing AGI Forecasts Forward

Epoch AI ranks GPT-6 Astra first overall with 169 points across 50 benchmarks, while Artificial Analysis scores it at 61, trailing Anthropic's Claude Fable 5.1. Astra scored 62.7% on ARC-AGI-3, a massive jump from GPT-5.6 Sol's 7.78% and Claude Opus 5's 30.16%. The model exhibits an inverted compute-cost curve, where applying maximum reasoning drops the test cost from $49,791 to $26,098 by requiring fewer moves to solve games. Astra cleared 96% of the tested levels in fewer moves than the median human baseline, effectively operating in the human efficiency range once it grasps the environment's mechanics. To manage complex states, the model autonomously generated its own algebraic domain-specific language (DSL) to compress and record coordinates and rules on the fly. "I believe the pace of progress will surprise many people, and what the new models are capable of will challenge the perception of AI that people have formed based on earlier generations of models." — François Chollet