GPT-6 ASTRA DOMINATES ARC-AGI BENCHMARK

OpenAI's GPT-6 Astra demonstrated substantial performance gains on the ARC-AGI-3 reasoning benchmark. Using the standard harness and $26K in compute, the model scored 62.7%—more than double Claude Opus 5's 30.2% and nearly eight times GPT-5.6 Sol's 7.8%. The results shift considerably with implementation changes. A provider adapter harness reduced costs to $19K while pushing Astra's score to 99.9%, suggesting optimization approaches can significantly impact model performance on abstract reasoning tasks. The ARC-AGI benchmark tests artificial general intelligence capabilities through pattern recognition and abstract reasoning problems. These latest results, reported by Greg Kamradt from the ARC Prize, indicate GPT-6 Astra represents a notable advancement in reasoning-focused AI performance. The gap between standard and optimized harness results raises questions about evaluation methodology consistency in AI benchmarking.